forked from M-Labs/nac3
1
0
Fork 0
nac3/nac3core/irrt/irrt_numpy_ndarray.hpp

481 lines
19 KiB
C++

#pragma once
#include "irrt_utils.hpp"
#include "irrt_typedefs.hpp"
#include "irrt_slice.hpp"
/*
NDArray-related implementations.
`*/
namespace {
namespace ndarray_util {
template <typename SizeT>
static void set_indices_by_nth(SizeT ndims, const SizeT* shape, SizeT* indices, SizeT nth) {
for (int32_t i = 0; i < ndims; i++) {
int32_t dim_i = ndims - i - 1;
int32_t dim = shape[dim_i];
indices[dim_i] = nth % dim;
nth /= dim;
}
}
// Compute the strides of an ndarray given an ndarray `shape`
// and assuming that the ndarray is *fully C-contagious*.
//
// You might want to read up on https://ajcr.net/stride-guide-part-1/.
template <typename SizeT>
static void set_strides_by_shape(SizeT itemsize, SizeT ndims, SizeT* dst_strides, const SizeT* shape) {
SizeT stride_product = 1;
for (SizeT i = 0; i < ndims; i++) {
int dim_i = ndims - i - 1;
dst_strides[dim_i] = stride_product * itemsize;
stride_product *= shape[dim_i];
}
}
// Compute the size/# of elements of an ndarray given its shape
template <typename SizeT>
static SizeT calc_size_from_shape(SizeT ndims, const SizeT* shape) {
SizeT size = 1;
for (SizeT dim_i = 0; dim_i < ndims; dim_i++) size *= shape[dim_i];
return size;
}
template <typename SizeT>
static bool can_broadcast_shape_to(
const SizeT target_ndims,
const SizeT *target_shape,
const SizeT src_ndims,
const SizeT *src_shape
) {
/*
// See https://numpy.org/doc/stable/user/basics.broadcasting.html
This function handles this example:
```
Image (3d array): 256 x 256 x 3
Scale (1d array): 3
Result (3d array): 256 x 256 x 3
```
Other interesting examples to consider:
- `can_broadcast_shape_to([3], [1, 1, 1, 1, 3]) == true`
- `can_broadcast_shape_to([3], [3, 1]) == false`
- `can_broadcast_shape_to([256, 256, 3], [256, 1, 3]) == true`
In cases when the shapes contain zero(es):
- `can_broadcast_shape_to([0], [1]) == true`
- `can_broadcast_shape_to([0], [2]) == false`
- `can_broadcast_shape_to([0, 4, 0, 0], [1]) == true`
- `can_broadcast_shape_to([0, 4, 0, 0], [1, 1, 1, 1]) == true`
- `can_broadcast_shape_to([0, 4, 0, 0], [1, 4, 1, 1]) == true`
- `can_broadcast_shape_to([4, 3], [0, 3]) == false`
- `can_broadcast_shape_to([4, 3], [0, 0]) == false`
*/
// This is essentially doing the following in Python:
// `for target_dim, src_dim in itertools.zip_longest(target_shape[::-1], src_shape[::-1], fillvalue=1)`
for (SizeT i = 0; i < max(target_ndims, src_ndims); i++) {
SizeT target_dim_i = target_ndims - i - 1;
SizeT src_dim_i = src_ndims - i - 1;
bool target_dim_exists = target_dim_i >= 0;
bool src_dim_exists = src_dim_i >= 0;
SizeT target_dim = target_dim_exists ? target_shape[target_dim_i] : 1;
SizeT src_dim = src_dim_exists ? src_shape[src_dim_i] : 1;
bool ok = src_dim == 1 || target_dim == src_dim;
if (!ok) return false;
}
return true;
}
}
typedef uint8_t NDSliceType;
extern "C" {
const NDSliceType INPUT_SLICE_TYPE_INDEX = 0;
const NDSliceType INPUT_SLICE_TYPE_SLICE = 1;
}
struct NDSlice {
// A poor-man's enum variant type
NDSliceType type;
/*
if type == INPUT_SLICE_TYPE_INDEX => `slice` points to a single `SizeT`
if type == INPUT_SLICE_TYPE_SLICE => `slice` points to a single `UserRange<SizeT>`
`SizeT` is controlled by the caller: `NDSlice` only cares about where that
slice is (the pointer), `NDSlice` does not care/know about the actual `sizeof()`
of the slice value.
*/
uint8_t* slice;
};
namespace ndarray_util {
template<typename SizeT>
SizeT deduce_ndims_after_slicing(SizeT ndims, SizeT num_slices, const NDSlice *slices) {
irrt_assert(num_slices <= ndims);
SizeT final_ndims = ndims;
for (SizeT i = 0; i < num_slices; i++) {
if (slices[i].type == INPUT_SLICE_TYPE_INDEX) {
final_ndims--; // An index demotes the rank by 1
}
}
return final_ndims;
}
}
// template <typename SizeT>
// struct NDArrayIndicesIter {
// SizeT ndims;
// const SizeT *shape;
// SizeT *indices;
// void set_indices_zero() {
// __builtin_memset(indices, 0, sizeof(SizeT) * ndims);
// }
// void next() {
// for (SizeT i = 0; i < ndims; i++) {
// SizeT dim_i = ndims - i - 1;
// indices[dim_i]++;
// if (indices[dim_i] < shape[dim_i]) {
// break;
// } else {
// indices[dim_i] = 0;
// }
// }
// }
// };
// The NDArray object. `SizeT` is the *signed* size type of this ndarray.
//
// NOTE: The order of fields is IMPORTANT. DON'T TOUCH IT
//
// Some resources you might find helpful:
// - The official numpy implementations:
// - https://github.com/numpy/numpy/blob/735a477f0bc2b5b84d0e72d92f224bde78d4e069/doc/source/reference/c-api/types-and-structures.rst
// - On strides (about reshaping, slicing, C-contagiousness, etc)
// - https://ajcr.net/stride-guide-part-1/.
// - https://ajcr.net/stride-guide-part-2/.
// - https://ajcr.net/stride-guide-part-3/.
template <typename SizeT>
struct NDArray {
// The underlying data this `ndarray` is pointing to.
//
// NOTE: Formally this should be of type `void *`, but clang
// translates `void *` to `i8 *` when run with `-S -emit-llvm`,
// so we will put `uint8_t *` here for clarity.
uint8_t *data;
// The number of bytes of a single element in `data`.
//
// The `SizeT` is treated as `unsigned`.
SizeT itemsize;
// The number of dimensions of this shape.
//
// The `SizeT` is treated as `unsigned`.
SizeT ndims;
// Array shape, with length equal to `ndims`.
//
// The `SizeT` is treated as `unsigned`.
//
// NOTE: `shape` can contain 0.
// (those appear when the user makes an out of bounds slice into an ndarray, e.g., `np.zeros((3, 3))[400:].shape == (0, 3)`)
SizeT *shape;
// Array strides (stride value is in number of bytes, NOT number of elements), with length equal to `ndims`.
//
// The `SizeT` is treated as `signed`.
//
// NOTE: `strides` can have negative numbers.
// (those appear when there is a slice with a negative step, e.g., `my_array[::-1]`)
SizeT *strides;
// Calculate the size/# of elements of an `ndarray`.
// This function corresponds to `np.size(<ndarray>)` or `ndarray.size`
SizeT size() {
return ndarray_util::calc_size_from_shape(ndims, shape);
}
// Calculate the number of bytes of its content of an `ndarray` *in its view*.
// This function corresponds to `ndarray.nbytes`
SizeT nbytes() {
return this->size() * itemsize;
}
void set_pelement_value(uint8_t* pelement, const uint8_t* pvalue) {
__builtin_memcpy(pelement, pvalue, itemsize);
}
uint8_t* get_pelement_by_indices(const SizeT *indices) {
uint8_t* element = data;
for (SizeT dim_i = 0; dim_i < ndims; dim_i++)
element += indices[dim_i] * strides[dim_i];
return element;
}
uint8_t* get_nth_pelement(SizeT nth) {
irrt_assert(0 <= nth);
irrt_assert(nth < this->size());
SizeT* indices = (SizeT*) __builtin_alloca(sizeof(SizeT) * this->ndims);
ndarray_util::set_indices_by_nth(this->ndims, this->shape, indices, nth);
return get_pelement_by_indices(indices);
}
// Get pointer to the first element of this ndarray, assuming
// `this->size() > 0`, i.e., not "degenerate" due to zeroes in `this->shape`)
//
// This is particularly useful for when the ndarray is just containing a single scalar.
uint8_t* get_first_pelement() {
irrt_assert(this->size() > 0);
return this->data; // ...It is simply `this->data`
}
// Is the given `indices` valid/in-bounds?
bool in_bounds(const SizeT *indices) {
for (SizeT dim_i = 0; dim_i < ndims; dim_i++) {
bool dim_ok = indices[dim_i] < shape[dim_i];
if (!dim_ok) return false;
}
return true;
}
// 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);
set_pelement_value(pelement, pvalue);
}
}
// Set the strides of the ndarray with `ndarray_util::set_strides_by_shape`
void set_strides_by_shape() {
ndarray_util::set_strides_by_shape(itemsize, ndims, strides, shape);
}
// https://numpy.org/doc/stable/reference/generated/numpy.eye.html
void set_to_eye(SizeT k, const uint8_t* zero_pvalue, const uint8_t* one_pvalue) {
__builtin_assume(ndims == 2);
// TODO: Better implementation
fill_generic(zero_pvalue);
for (SizeT i = 0; i < min(shape[0], shape[1]); i++) {
SizeT row = i;
SizeT col = i + k;
SizeT indices[2] = { row, col };
if (!in_bounds(indices)) continue;
uint8_t* pelement = get_pelement_by_indices(indices);
set_pelement_value(pelement, one_pvalue);
}
}
// To support numpy "basic indexing" https://numpy.org/doc/stable/user/basics.indexing.html#basic-indexing
// "Advanced indexing" https://numpy.org/doc/stable/user/basics.indexing.html#advanced-indexing is not supported
//
// This function supports:
// - "scalar indexing",
// - "slicing and strides",
// - and "dimensional indexing tools" (TODO, but this is really easy to implement).
//
// Things assumed by this function:
// - `dst_ndarray` is allocated by the caller
// - `dst_ndarray.ndims` has the correct value (according to `ndarray_util::deduce_ndims_after_slicing`).
// - ... and `dst_ndarray.shape` and `dst_ndarray.strides` have been allocated by the caller as well
//
// Other notes:
// - `dst_ndarray->data` does not have to be set, it will be derived.
// - `dst_ndarray->itemsize` does not have to be set, it will be set to `this->itemsize`
// - `dst_ndarray->shape` and `dst_ndarray.strides` can contain empty values
void subscript(SizeT num_ndslices, NDSlice* ndslices, NDArray<SizeT>* dst_ndarray) {
// REFERENCE CODE (check out `_index_helper` in `__getitem__`):
// https://github.com/wadetb/tinynumpy/blob/0d23d22e07062ffab2afa287374c7b366eebdda1/tinynumpy/tinynumpy.py#L652
irrt_assert(dst_ndarray->ndims == ndarray_util::deduce_ndims_after_slicing(this->ndims, num_ndslices, ndslices));
dst_ndarray->data = this->data;
SizeT this_axis = 0;
SizeT dst_axis = 0;
for (SizeT i = 0; i < num_ndslices; i++) {
NDSlice *ndslice = &ndslices[i];
if (ndslice->type == INPUT_SLICE_TYPE_INDEX) {
// Handle when the ndslice is just a single (possibly negative) integer
// e.g., `my_array[::2, -5, ::-1]`
// ^^------ like this
SizeT index_user = *((SizeT*) ndslice->slice);
SizeT index = resolve_index_in_length(this->shape[this_axis], index_user);
dst_ndarray->data += index * this->strides[this_axis]; // Add offset
// Next
this_axis++;
} else if (ndslice->type == INPUT_SLICE_TYPE_SLICE) {
// Handle when the ndslice is a slice (represented by UserSlice in IRRT)
// e.g., `my_array[::2, -5, ::-1]`
// ^^^------^^^^----- like these
UserSlice* user_slice = (UserSlice*) ndslice->slice;
Slice slice = user_slice->indices(this->shape[this_axis]); // To resolve negative indices and other funny stuff written by the user
// NOTE: There is no need to write special code to handle negative steps/strides.
// This simple implementation meticulously handles both positive and negative steps/strides.
// Check out the tinynumpy and IRRT's test cases if you are not convinced.
dst_ndarray->data += (SizeT) slice.start * this->strides[this_axis]; // Add offset (NOTE: no need to `* itemsize`, strides count in # of bytes)
dst_ndarray->strides[dst_axis] = ((SizeT) slice.step) * this->strides[this_axis]; // Determine stride
dst_ndarray->shape[dst_axis] = (SizeT) slice.len(); // Determine shape dimension
// Next
dst_axis++;
this_axis++;
} else {
__builtin_unreachable();
}
}
irrt_assert(dst_axis == dst_ndarray->ndims); // Sanity check on the implementation
}
// Similar to `np.broadcast_to(<ndarray>, <target_shape>)`
// Assumptions:
// - `this` has to be fully initialized.
// - `dst_ndarray->ndims` has to be set.
// - `dst_ndarray->shape` has to be set, this determines the shape `this` broadcasts to.
//
// Other notes:
// - `dst_ndarray->data` does not have to be set, it will be set to `this->data`.
// - `dst_ndarray->itemsize` does not have to be set, it will be set to `this->data`.
// - `dst_ndarray->strides` does not have to be set, it will be overwritten.
//
// Cautions:
// ```
// xs = np.zeros((4,))
// ys = np.zero((4, 1))
// ys[:] = xs # ok
//
// xs = np.zeros((1, 4))
// ys = np.zero((4,))
// ys[:] = xs # allowed
// # However `np.broadcast_to(xs, (4,))` would fails, as per numpy's broadcasting rule.
// # and apparently numpy will "deprecate" this? SEE https://github.com/numpy/numpy/issues/21744
// # This implementation will NOT support this assignment.
// ```
void broadcast_to(NDArray<SizeT>* dst_ndarray) {
dst_ndarray->data = this->data;
dst_ndarray->itemsize = this->itemsize;
irrt_assert(
ndarray_util::can_broadcast_shape_to(
dst_ndarray->ndims,
dst_ndarray->shape,
this->ndims,
this->shape
)
);
SizeT stride_product = 1;
for (SizeT i = 0; i < max(this->ndims, dst_ndarray->ndims); i++) {
SizeT this_dim_i = this->ndims - i - 1;
SizeT dst_dim_i = dst_ndarray->ndims - i - 1;
bool this_dim_exists = this_dim_i >= 0;
bool dst_dim_exists = dst_dim_i >= 0;
// TODO: Explain how this works
bool c1 = this_dim_exists && this->shape[this_dim_i] == 1;
bool c2 = dst_dim_exists && dst_ndarray->shape[dst_dim_i] != 1;
if (!this_dim_exists || (c1 && c2)) {
dst_ndarray->strides[dst_dim_i] = 0; // Freeze it in-place
} else {
dst_ndarray->strides[dst_dim_i] = stride_product * this->itemsize;
stride_product *= this->shape[this_dim_i]; // NOTE: this_dim_exist must be true here.
}
}
}
// Simulates `this_ndarray[:] = src_ndarray`, with automatic broadcasting.
// Caution on https://github.com/numpy/numpy/issues/21744
// Also see `NDArray::broadcast_to`
void assign_with(NDArray<SizeT>* src_ndarray) {
irrt_assert(
ndarray_util::can_broadcast_shape_to(
this->ndims,
this->shape,
src_ndarray->ndims,
src_ndarray->shape
)
);
// Broadcast the `src_ndarray` to make the reading process *much* easier
SizeT* broadcasted_src_ndarray_strides = __builtin_alloca(sizeof(SizeT) * this->ndims); // Remember to allocate strides beforehand
NDArray<SizeT> broadcasted_src_ndarray = {
.ndims = this->ndims,
.shape = this->shape,
.strides = broadcasted_src_ndarray_strides
};
src_ndarray->broadcast_to(&broadcasted_src_ndarray);
const SizeT size = this->size();
for (SizeT i = 0; i < size; i++) {
uint8_t* src_pelement = broadcasted_src_ndarray_strides->get_nth_pelement(i);
uint8_t* this_pelement = this->get_nth_pelement(i);
this->set_pelement_value(this_pelement, src_pelement);
}
}
};
}
extern "C" {
uint32_t __nac3_ndarray_size(NDArray<int32_t>* ndarray) {
return ndarray->size();
}
uint64_t __nac3_ndarray_size64(NDArray<int64_t>* ndarray) {
return ndarray->size();
}
void __nac3_ndarray_set_strides_by_shape(NDArray<int32_t>* ndarray) {
ndarray->set_strides_by_shape();
}
void __nac3_ndarray_set_strides_by_shape64(NDArray<int64_t>* ndarray) {
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);
}
int32_t __nac3_ndarray_deduce_ndims_after_slicing(int32_t ndims, int32_t num_slices, const NDSlice* slices) {
return ndarray_util::deduce_ndims_after_slicing(ndims, num_slices, slices);
}
int64_t __nac3_ndarray_deduce_ndims_after_slicing64(int64_t ndims, int64_t num_slices, const NDSlice* slices) {
return ndarray_util::deduce_ndims_after_slicing(ndims, num_slices, slices);
}
void __nac3_ndarray_subscript(NDArray<int32_t>* ndarray, int32_t num_slices, NDSlice* slices, NDArray<int32_t> *dst_ndarray) {
ndarray->subscript(num_slices, slices, dst_ndarray);
}
void __nac3_ndarray_subscript64(NDArray<int64_t>* ndarray, int32_t num_slices, NDSlice* slices, NDArray<int64_t> *dst_ndarray) {
ndarray->subscript(num_slices, slices, dst_ndarray);
}
}