forked from M-Labs/nac3
1
0
Fork 0

core: ndarray fill generic

This commit is contained in:
lyken 2024-07-14 15:45:06 +08:00
parent 3b87bd36f3
commit 9a82b033b6
4 changed files with 94 additions and 66 deletions

View File

@ -96,6 +96,19 @@ struct NDArray {
} }
return get_nth_pelement(nth); return get_nth_pelement(nth);
} }
void set_pelement_value(uint8_t* pelement, const uint8_t* pvalue) {
__builtin_memcpy(pelement, pvalue, itemsize);
}
// Fill the ndarray with a value
void fill_generic(const uint8_t* pvalue) {
const SizeT size = this->size();
for (SizeT i = 0; i < size; i++) {
uint8_t* pelement = get_nth_pelement(i); // No need for checked_get_nth_pelement
set_pelement_value(pelement, pvalue);
}
}
}; };
} }
@ -131,4 +144,12 @@ void __nac3_ndarray_set_strides_by_shape(NDArray<int32_t>* ndarray) {
void __nac3_ndarray_set_strides_by_shape64(NDArray<int64_t>* ndarray) { void __nac3_ndarray_set_strides_by_shape64(NDArray<int64_t>* ndarray) {
ndarray->set_strides_by_shape(); ndarray->set_strides_by_shape();
} }
void __nac3_ndarray_fill_generic(NDArray<int32_t>* ndarray, uint8_t* pvalue) {
ndarray->fill_generic(pvalue);
}
void __nac3_ndarray_fill_generic64(NDArray<int64_t>* ndarray, uint8_t* pvalue) {
ndarray->fill_generic(pvalue);
}
} }

View File

@ -1,9 +1,7 @@
use inkwell::types::IntType; use inkwell::types::IntType;
use crate::codegen::{ use crate::codegen::{
optics::{ optics::{Address, AddressLens, ArraySlice, FieldBuilder, GepGetter, IntLens, StructureOptic},
Address, AddressLens, ArraySlice, FieldBuilder, GepGetter, IntLens, Optic, StructureOptic,
},
CodeGenContext, CodeGenContext,
}; };
@ -41,18 +39,12 @@ pub struct NpArrayFields<'ctx> {
pub strides: GepGetter<AddressLens<IntLens<'ctx>>>, pub strides: GepGetter<AddressLens<IntLens<'ctx>>>,
} }
// Note: NpArrayLens's ElementOptic is purely for type-safety and type-guidances
// The underlying LLVM ndarray doesn't care, it only holds an opaque (uint8_t*) pointer to the elements.
#[derive(Debug, Clone, Copy)] #[derive(Debug, Clone, Copy)]
pub struct NpArrayLens<'ctx, ElementOptic> { pub struct NpArrayLens<'ctx> {
pub size_type: IntType<'ctx>, pub size_type: IntType<'ctx>,
pub element_optic: ElementOptic,
} }
// NDArray is *frequently* used, so here is a type alias impl<'ctx> StructureOptic<'ctx> for NpArrayLens<'ctx> {
pub type NpArray<'ctx, ElementOptic> = Address<'ctx, NpArrayLens<'ctx, ElementOptic>>;
impl<'ctx, ElementOptic: Optic<'ctx>> StructureOptic<'ctx> for NpArrayLens<'ctx, ElementOptic> {
type Fields = NpArrayFields<'ctx>; type Fields = NpArrayFields<'ctx>;
fn struct_name(&self) -> &'static str { fn struct_name(&self) -> &'static str {
@ -74,7 +66,7 @@ impl<'ctx, ElementOptic: Optic<'ctx>> StructureOptic<'ctx> for NpArrayLens<'ctx,
} }
// Other convenient utilities for NpArray // Other convenient utilities for NpArray
impl<'ctx, ElementOptic: Optic<'ctx>> NpArray<'ctx, ElementOptic> { impl<'ctx> Address<'ctx, NpArrayLens<'ctx>> {
pub fn shape_array(&self, ctx: &CodeGenContext<'ctx, '_>) -> ArraySlice<'ctx, IntLens<'ctx>> { pub fn shape_array(&self, ctx: &CodeGenContext<'ctx, '_>) -> ArraySlice<'ctx, IntLens<'ctx>> {
let ndims = self.focus(ctx, |fields| &fields.ndims).load(ctx, "ndims"); let ndims = self.focus(ctx, |fields| &fields.ndims).load(ctx, "ndims");
let shape_base_ptr = self.focus(ctx, |fields| &fields.shape).load(ctx, "shape"); let shape_base_ptr = self.focus(ctx, |fields| &fields.shape).load(ctx, "shape");

View File

@ -1,14 +1,14 @@
use std::marker::PhantomData; use std::marker::PhantomData;
use inkwell::{ use inkwell::{
types::BasicType, types::{BasicType, BasicTypeEnum},
values::{BasicValueEnum, IntValue}, values::{BasicValueEnum, IntValue},
}; };
use crate::{ use crate::{
codegen::{ codegen::{
classes::{ListValue, UntypedArrayLikeAccessor}, classes::{ListValue, UntypedArrayLikeAccessor},
optics::{Address, AddressLens, ArraySlice, IntLens, Ixed, Optic}, optics::{opaque_address_lens, Address, AddressLens, ArraySlice, IntLens, Ixed, Optic},
stmt::gen_for_callback_incrementing, stmt::gen_for_callback_incrementing,
CodeGenContext, CodeGenerator, CodeGenContext, CodeGenerator,
}, },
@ -16,7 +16,7 @@ use crate::{
}; };
use super::{ use super::{
classes::{ErrorContextLens, NpArray, NpArrayLens}, classes::{ErrorContextLens, NpArrayLens},
new::{ new::{
check_error_context, get_sized_dependent_function_name, prepare_error_context, check_error_context, get_sized_dependent_function_name, prepare_error_context,
FunctionBuilder, FunctionBuilder,
@ -176,38 +176,43 @@ where
} }
} }
pub fn alloca_ndarray<'ctx, G, ElementOptic: Optic<'ctx>>( pub fn alloca_ndarray<'ctx, G>(
generator: &mut G, generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>, ctx: &mut CodeGenContext<'ctx, '_>,
element_optic: ElementOptic, elem_type: BasicTypeEnum<'ctx>,
ndims: IntValue<'ctx>, ndims: IntValue<'ctx>,
name: &str, name: &str,
) -> Result<NpArray<'ctx, ElementOptic>, String> ) -> Result<Address<'ctx, NpArrayLens<'ctx>>, String>
where where
G: CodeGenerator + ?Sized, G: CodeGenerator + ?Sized,
{ {
let size_type = generator.get_size_type(ctx.ctx); let size_type = generator.get_size_type(ctx.ctx);
let itemsize = element_optic.get_llvm_type(ctx.ctx).size_of().unwrap(); // Allocate ndarray
let ndarray_ptr = NpArrayLens { size_type }.alloca(ctx, name);
// Set ndims
ndarray_ptr.focus(ctx, |fields| &fields.ndims).store(ctx, &ndims);
// Set itemsize
let itemsize = elem_type.size_of().unwrap();
let itemsize = let itemsize =
ctx.builder.build_int_s_extend_or_bit_cast(itemsize, size_type, "itemsize").unwrap(); ctx.builder.build_int_s_extend_or_bit_cast(itemsize, size_type, "itemsize").unwrap();
ndarray_ptr.focus(ctx, |fields| &fields.itemsize).store(ctx, &itemsize);
let shape = ctx.builder.build_array_alloca(size_type, ndims, "shape").unwrap(); // Allocate and set shape
let strides = ctx.builder.build_array_alloca(size_type, ndims, "strides").unwrap(); let shape_ptr = ctx.builder.build_array_alloca(size_type, ndims, "shape").unwrap();
ndarray_ptr
let ndarray = NpArrayLens { size_type, element_optic }.alloca(ctx, name);
// Set ndims, itemsize; and allocate shape and store on the stack
ndarray.focus(ctx, |fields| &fields.ndims).store(ctx, &ndims);
ndarray.focus(ctx, |fields| &fields.itemsize).store(ctx, &itemsize);
ndarray
.focus(ctx, |fields| &fields.shape) .focus(ctx, |fields| &fields.shape)
.store(ctx, &Address { addressee_optic: IntLens(size_type), address: shape }); .store(ctx, &Address { addressee_optic: IntLens(size_type), address: shape_ptr });
ndarray
.focus(ctx, |fields| &fields.strides)
.store(ctx, &Address { addressee_optic: IntLens(size_type), address: strides });
Ok(ndarray) // Allocate and set strides
let strides_ptr = ctx.builder.build_array_alloca(size_type, ndims, "strides").unwrap();
ndarray_ptr
.focus(ctx, |fields| &fields.strides)
.store(ctx, &Address { addressee_optic: IntLens(size_type), address: strides_ptr });
Ok(ndarray_ptr)
} }
enum NDArrayInitMode<'ctx, G: CodeGenerator + ?Sized> { enum NDArrayInitMode<'ctx, G: CodeGenerator + ?Sized> {
@ -217,75 +222,75 @@ enum NDArrayInitMode<'ctx, G: CodeGenerator + ?Sized> {
} }
/// TODO: DOCUMENT ME /// TODO: DOCUMENT ME
fn alloca_ndarray_and_init<'ctx, G, ElementOptic: Optic<'ctx>>( fn alloca_ndarray_and_init<'ctx, G>(
generator: &mut G, generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>, ctx: &mut CodeGenContext<'ctx, '_>,
element_optic: ElementOptic, elem_type: BasicTypeEnum<'ctx>,
init_mode: NDArrayInitMode<'ctx, G>, init_mode: NDArrayInitMode<'ctx, G>,
name: &str, name: &str,
) -> Result<NpArray<'ctx, ElementOptic>, String> ) -> Result<Address<'ctx, NpArrayLens<'ctx>>, String>
where where
G: CodeGenerator + ?Sized, G: CodeGenerator + ?Sized,
{ {
// It is implemented verbosely in order to make the initialization modes super clear in their intent. // It is implemented verbosely in order to make the initialization modes super clear in their intent.
match init_mode { match init_mode {
NDArrayInitMode::NDim { ndim: ndims, _phantom } => { NDArrayInitMode::NDim { ndim: ndims, _phantom } => {
let ndarray = alloca_ndarray(generator, ctx, element_optic, ndims, name)?; let ndarray = alloca_ndarray(generator, ctx, elem_type, ndims, name)?;
Ok(ndarray) Ok(ndarray)
} }
NDArrayInitMode::Shape { shape } => { NDArrayInitMode::Shape { shape } => {
let ndims = shape.count; let ndims = shape.count;
let ndarray = alloca_ndarray(generator, ctx, element_optic, ndims, name)?; let ndarray_ptr = alloca_ndarray(generator, ctx, elem_type, ndims, name)?;
// Fill `ndarray.shape` // Fill `ndarray.shape`
(shape.write_to_array)(generator, ctx, &ndarray.shape_array(ctx))?; (shape.write_to_array)(generator, ctx, &ndarray_ptr.shape_array(ctx))?;
// Check if `shape` has bad inputs // Check if `shape` has bad inputs
call_nac3_ndarray_util_assert_shape_no_negative( call_nac3_ndarray_util_assert_shape_no_negative(
generator, generator,
ctx, ctx,
ndims, ndims,
&ndarray.focus(ctx, |fields| &fields.shape).load(ctx, "shape"), &ndarray_ptr.focus(ctx, |fields| &fields.shape).load(ctx, "shape"),
); );
// NOTE: DO NOT DO `set_strides_by_shape` HERE. // NOTE: DO NOT DO `set_strides_by_shape` HERE.
// Simply this is because we specified that `SetShape` wouldn't do `set_strides_by_shape` // Simply this is because we specified that `SetShape` wouldn't do `set_strides_by_shape`
Ok(ndarray) Ok(ndarray_ptr)
} }
NDArrayInitMode::ShapeAndAllocaData { shape } => { NDArrayInitMode::ShapeAndAllocaData { shape } => {
let ndims = shape.count; let ndims = shape.count;
let ndarray = alloca_ndarray(generator, ctx, element_optic, ndims, name)?; let ndarray_ptr = alloca_ndarray(generator, ctx, elem_type, ndims, name)?;
// Fill `ndarray.shape` // Fill `ndarray.shape`
(shape.write_to_array)(generator, ctx, &ndarray.shape_array(ctx))?; (shape.write_to_array)(generator, ctx, &ndarray_ptr.shape_array(ctx))?;
// Check if `shape` has bad inputs // Check if `shape` has bad inputs
call_nac3_ndarray_util_assert_shape_no_negative( call_nac3_ndarray_util_assert_shape_no_negative(
generator, generator,
ctx, ctx,
ndims, ndims,
&ndarray.focus(ctx, |fields| &fields.shape).load(ctx, "shape"), &ndarray_ptr.focus(ctx, |fields| &fields.shape).load(ctx, "shape"),
); );
// Now we populate `ndarray.data` by alloca-ing. // Now we populate `ndarray.data` by alloca-ing.
// But first, we need to know the size of the ndarray to know how many elements to alloca, // But first, we need to know the size of the ndarray to know how many elements to alloca,
// since calculating nbytes of an ndarray requires `ndarray.shape` to be set. // since calculating nbytes of an ndarray requires `ndarray.shape` to be set.
let ndarray_nbytes = call_nac3_ndarray_nbytes(generator, ctx, &ndarray); let ndarray_nbytes = call_nac3_ndarray_nbytes(generator, ctx, &ndarray_ptr);
// Alloca `data` and assign it to `ndarray.data` // Alloca `data` and assign it to `ndarray.data`
let data_ptr = let data_ptr =
ctx.builder.build_array_alloca(ctx.ctx.i8_type(), ndarray_nbytes, "data").unwrap(); ctx.builder.build_array_alloca(ctx.ctx.i8_type(), ndarray_nbytes, "data").unwrap();
ndarray.focus(ctx, |fields| &fields.data).store( ndarray_ptr.focus(ctx, |fields| &fields.data).store(
ctx, ctx,
&Address { addressee_optic: IntLens::int8(ctx.ctx), address: data_ptr }, &Address { addressee_optic: IntLens::int8(ctx.ctx), address: data_ptr },
); );
// Finally, do `set_strides_by_shape` // Finally, do `set_strides_by_shape`
// Check out https://ajcr.net/stride-guide-part-1/ to see what numpy "strides" are. // Check out https://ajcr.net/stride-guide-part-1/ to see what numpy "strides" are.
call_nac3_ndarray_set_strides_by_shape(generator, ctx, &ndarray); call_nac3_ndarray_set_strides_by_shape(generator, ctx, &ndarray_ptr);
Ok(ndarray) Ok(ndarray_ptr)
} }
} }
} }
@ -294,7 +299,7 @@ fn call_nac3_ndarray_util_assert_shape_no_negative<'ctx, G: CodeGenerator + ?Siz
generator: &mut G, generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>, ctx: &mut CodeGenContext<'ctx, '_>,
ndims: IntValue<'ctx>, ndims: IntValue<'ctx>,
shape: &Address<'ctx, IntLens<'ctx>>, shape_ptr: &Address<'ctx, IntLens<'ctx>>,
) { ) {
let size_type = generator.get_size_type(ctx.ctx); let size_type = generator.get_size_type(ctx.ctx);
@ -308,19 +313,15 @@ fn call_nac3_ndarray_util_assert_shape_no_negative<'ctx, G: CodeGenerator + ?Siz
) )
.arg("errctx", &AddressLens(ErrorContextLens), &errctx) .arg("errctx", &AddressLens(ErrorContextLens), &errctx)
.arg("ndims", &IntLens(size_type), &ndims) .arg("ndims", &IntLens(size_type), &ndims)
.arg("shape", &AddressLens(IntLens(size_type)), shape) .arg("shape", &AddressLens(IntLens(size_type)), shape_ptr)
.returning_void(); .returning_void();
check_error_context(generator, ctx, &errctx); check_error_context(generator, ctx, &errctx);
} }
fn call_nac3_ndarray_set_strides_by_shape< fn call_nac3_ndarray_set_strides_by_shape<'ctx, G: CodeGenerator + ?Sized>(
'ctx,
G: CodeGenerator + ?Sized,
ElementOptic: Optic<'ctx>,
>(
generator: &mut G, generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>, ctx: &mut CodeGenContext<'ctx, '_>,
ndarray: &NpArray<'ctx, ElementOptic>, ndarray_ptr: &Address<'ctx, NpArrayLens<'ctx>>,
) { ) {
let size_type = generator.get_size_type(ctx.ctx); let size_type = generator.get_size_type(ctx.ctx);
@ -331,14 +332,14 @@ fn call_nac3_ndarray_set_strides_by_shape<
"__nac3_ndarray_util_assert_shape_no_negative", "__nac3_ndarray_util_assert_shape_no_negative",
), ),
) )
.arg("ndarray", &AddressLens(ndarray.addressee_optic.clone()), ndarray) .arg("ndarray", &AddressLens(NpArrayLens { size_type }), ndarray_ptr)
.returning_void(); .returning_void();
} }
fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized, ElementOptic: Optic<'ctx>>( fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G, generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>, ctx: &mut CodeGenContext<'ctx, '_>,
ndarray: &NpArray<'ctx, ElementOptic>, ndarray_ptr: &Address<'ctx, NpArrayLens<'ctx>>,
) -> IntValue<'ctx> { ) -> IntValue<'ctx> {
let size_type = generator.get_size_type(ctx.ctx); let size_type = generator.get_size_type(ctx.ctx);
@ -349,6 +350,23 @@ fn call_nac3_ndarray_nbytes<'ctx, G: CodeGenerator + ?Sized, ElementOptic: Optic
"__nac3_ndarray_util_assert_shape_no_negative", "__nac3_ndarray_util_assert_shape_no_negative",
), ),
) )
.arg("ndarray", &AddressLens(ndarray.addressee_optic.clone()), ndarray) .arg("ndarray", &AddressLens(NpArrayLens { size_type }), ndarray_ptr)
.returning("nbytes", &IntLens(size_type)) .returning("nbytes", &IntLens(size_type))
} }
fn call_nac3_ndarray_fill_generic<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray_ptr: &Address<'ctx, NpArrayLens<'ctx>>,
fill_value_ptr: &Address<'ctx, IntLens<'ctx>>,
) {
let size_type = generator.get_size_type(ctx.ctx);
FunctionBuilder::begin(
ctx,
&get_sized_dependent_function_name(size_type, "__nac3_ndarray_fill_generic"),
)
.arg("ndarray", &AddressLens(NpArrayLens { size_type }), ndarray_ptr)
.arg("pvalue", &opaque_address_lens(ctx.ctx), fill_value_ptr)
.returning_void();
}

View File

@ -50,8 +50,6 @@ pub trait MemorySetter<'ctx>: Optic<'ctx> {
fn set(&self, ctx: &CodeGenContext<'ctx, '_>, pointer: PointerValue<'ctx>, value: &Self::Value); fn set(&self, ctx: &CodeGenContext<'ctx, '_>, pointer: PointerValue<'ctx>, value: &Self::Value);
} }
pub trait SizedIntLens<'ctx>: Optic<'ctx, Value = IntValue<'ctx>> {}
// NOTE: I wanted to make Int8Lens, Int16Lens, Int32Lens, with all // NOTE: I wanted to make Int8Lens, Int16Lens, Int32Lens, with all
// having the trait IsIntLens, and implement `impl <S: IsIntLens> Optic<S> for T`, // having the trait IsIntLens, and implement `impl <S: IsIntLens> Optic<S> for T`,
// but that clashes with StructureOptic!! // but that clashes with StructureOptic!!
@ -141,10 +139,9 @@ impl<'ctx, AddresseeOptic> OpticValue<'ctx> for Address<'ctx, AddresseeOptic> {
#[derive(Debug, Clone)] #[derive(Debug, Clone)]
pub struct AddressLens<AddresseeOptic>(pub AddresseeOptic); pub struct AddressLens<AddresseeOptic>(pub AddresseeOptic);
impl<AddresseeOptic> AddressLens<AddresseeOptic> { #[must_use]
pub fn new_opaque<'ctx>(&self, ctx: &CodeGenContext<'ctx, '_>) -> AddressLens<IntLens<'ctx>> { pub fn opaque_address_lens(ctx: &Context) -> AddressLens<IntLens<'_>> {
AddressLens(IntLens::int8(ctx.ctx)) AddressLens(IntLens::int8(ctx))
}
} }
impl<'ctx, AddresseeOptic: Optic<'ctx>> Optic<'ctx> for AddressLens<AddresseeOptic> { impl<'ctx, AddresseeOptic: Optic<'ctx>> Optic<'ctx> for AddressLens<AddresseeOptic> {