forked from M-Labs/nac3
1
0
Fork 0

core: irrt general numpy broadcasting

This commit is contained in:
lyken 2024-07-10 17:05:01 +08:00
parent d18c769cdc
commit 9aae290727
3 changed files with 364 additions and 14 deletions

View File

@ -13,6 +13,17 @@ using NDIndex = uint32_t;
namespace { namespace {
namespace ndarray_util { 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` // Compute the strides of an ndarray given an ndarray `shape`
// and assuming that the ndarray is *fully C-contagious*. // and assuming that the ndarray is *fully C-contagious*.
// //
@ -34,6 +45,57 @@ namespace {
for (SizeT dim_i = 0; dim_i < ndims; dim_i++) size *= shape[dim_i]; for (SizeT dim_i = 0; dim_i < ndims; dim_i++) size *= shape[dim_i];
return size; 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; typedef uint8_t NDSliceType;
@ -55,7 +117,7 @@ namespace {
namespace ndarray_util { namespace ndarray_util {
template<typename SizeT> template<typename SizeT>
SizeT deduce_ndims_after_slicing(SizeT ndims, const SizeT num_slices, const NDSlice *slices) { SizeT deduce_ndims_after_slicing(SizeT ndims, SizeT num_slices, const NDSlice *slices) {
irrt_assert(num_slices <= ndims); irrt_assert(num_slices <= ndims);
SizeT final_ndims = ndims; SizeT final_ndims = ndims;
@ -150,17 +212,26 @@ namespace {
return this->size() * itemsize; return this->size() * itemsize;
} }
void set_value_at_pelement(uint8_t* pelement, uint8_t* pvalue) { void set_value_at_pelement(uint8_t* pelement, const uint8_t* pvalue) {
__builtin_memcpy(pelement, pvalue, itemsize); __builtin_memcpy(pelement, pvalue, itemsize);
} }
uint8_t* get_pelement(SizeT *indices) { uint8_t* get_pelement(const SizeT *indices) {
uint8_t* element = data; uint8_t* element = data;
for (SizeT dim_i = 0; dim_i < ndims; dim_i++) for (SizeT dim_i = 0; dim_i < ndims; dim_i++)
element += indices[dim_i] * strides[dim_i]; element += indices[dim_i] * strides[dim_i];
return element; 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(indices);
}
// Get pointer to the first element of this ndarray, assuming // Get pointer to the first element of this ndarray, assuming
// `this->size() > 0`, i.e., not "degenerate" due to zeroes in `this->shape`) // `this->size() > 0`, i.e., not "degenerate" due to zeroes in `this->shape`)
// //
@ -171,7 +242,7 @@ namespace {
} }
// Is the given `indices` valid/in-bounds? // Is the given `indices` valid/in-bounds?
bool in_bounds(SizeT *indices) { bool in_bounds(const SizeT *indices) {
for (SizeT dim_i = 0; dim_i < ndims; dim_i++) { for (SizeT dim_i = 0; dim_i < ndims; dim_i++) {
bool dim_ok = indices[dim_i] < shape[dim_i]; bool dim_ok = indices[dim_i] < shape[dim_i];
if (!dim_ok) return false; if (!dim_ok) return false;
@ -180,7 +251,7 @@ namespace {
} }
// Fill the ndarray with a value // Fill the ndarray with a value
void fill_generic(uint8_t* pvalue) { void fill_generic(const uint8_t* pvalue) {
NDArrayIndicesIter<SizeT> iter; NDArrayIndicesIter<SizeT> iter;
iter.ndims = this->ndims; iter.ndims = this->ndims;
iter.shape = this->shape; iter.shape = this->shape;
@ -199,7 +270,7 @@ namespace {
} }
// https://numpy.org/doc/stable/reference/generated/numpy.eye.html // https://numpy.org/doc/stable/reference/generated/numpy.eye.html
void set_to_eye(SizeT k, uint8_t* zero_pvalue, uint8_t* one_pvalue) { void set_to_eye(SizeT k, const uint8_t* zero_pvalue, const uint8_t* one_pvalue) {
__builtin_assume(ndims == 2); __builtin_assume(ndims == 2);
// TODO: Better implementation // TODO: Better implementation
@ -275,6 +346,63 @@ namespace {
irrt_assert(dst_axis == dst_ndarray->ndims); // Sanity check on the implementation 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.
}
}
}
}; };
} }

View File

@ -33,10 +33,11 @@ void debug_print_array(const char* format, int len, T* as) {
template <typename T> template <typename T>
void assert_arrays_match(const char* label, const char* format, int len, T* expected, T* got) { void assert_arrays_match(const char* label, const char* format, int len, T* expected, T* got) {
if (!arrays_match(len, expected, got)) { if (!arrays_match(len, expected, got)) {
printf("expected %s: ", label); printf(">>>>>>> %s\n", label);
printf(" Expecting = ");
debug_print_array(format, len, expected); debug_print_array(format, len, expected);
printf("\n"); printf("\n");
printf("got %s: ", label); printf(" Got = ");
debug_print_array(format, len, got); debug_print_array(format, len, got);
printf("\n"); printf("\n");
test_fail(); test_fail();
@ -46,22 +47,89 @@ void assert_arrays_match(const char* label, const char* format, int len, T* expe
template <typename T> template <typename T>
void assert_values_match(const char* label, const char* format, T expected, T got) { void assert_values_match(const char* label, const char* format, T expected, T got) {
if (expected != got) { if (expected != got) {
printf("expected %s: ", label); printf(">>>>>>> %s\n", label);
printf(" Expecting = ");
printf(format, expected); printf(format, expected);
printf("\n"); printf("\n");
printf("got %s: ", label); printf(" Got = ");
printf(format, got); printf(format, got);
printf("\n"); printf("\n");
test_fail(); test_fail();
} }
} }
void print_repeated(const char *str, int count) {
for (int i = 0; i < count; i++) {
printf("%s", str);
}
}
template<typename SizeT, typename ElementT>
void __print_ndarray_aux(const char *format, bool first, bool last, SizeT* cursor, SizeT depth, NDArray<SizeT>* ndarray) {
// A really lazy recursive implementation
// Add left padding unless its the first entry (since there would be "[[[" before it)
if (!first) {
print_repeated(" ", depth);
}
const SizeT dim = ndarray->shape[depth];
if (depth + 1 == ndarray->ndims) {
// Recursed down to last dimension, print the values in a nice list
printf("[");
SizeT* indices = (SizeT*) __builtin_alloca(sizeof(SizeT) * ndarray->ndims);
for (SizeT i = 0; i < dim; i++) {
ndarray_util::set_indices_by_nth(ndarray->ndims, ndarray->shape, indices, *cursor);
ElementT* pelement = (ElementT*) ndarray->get_pelement(indices);
ElementT element = *pelement;
if (i != 0) printf(", "); // List delimiter
printf(format, element);
printf("(@");
debug_print_array("%d", ndarray->ndims, indices);
printf(")");
(*cursor)++;
}
printf("]");
} else {
printf("[");
for (SizeT i = 0; i < ndarray->shape[depth]; i++) {
__print_ndarray_aux<SizeT, ElementT>(
format,
i == 0, // first?
i + 1 == dim, // last?
cursor,
depth + 1,
ndarray
);
}
printf("]");
}
// Add newline unless its the last entry (since there will be "]]]" after it)
if (!last) {
print_repeated("\n", depth);
}
}
template<typename SizeT, typename ElementT>
void print_ndarray(const char *format, NDArray<SizeT>* ndarray) {
if (ndarray->ndims == 0) {
printf("<empty ndarray>");
} else {
SizeT cursor = 0;
__print_ndarray_aux<SizeT, ElementT>(format, true, true, &cursor, 0, ndarray);
}
printf("\n");
}
void test_calc_size_from_shape_normal() { void test_calc_size_from_shape_normal() {
// Test shapes with normal values // Test shapes with normal values
BEGIN_TEST(); BEGIN_TEST();
int32_t shape[4] = { 2, 3, 5, 7 }; int32_t shape[4] = { 2, 3, 5, 7 };
debug_print_array("%d", 4, shape);
assert_values_match("size", "%d", 210, ndarray_util::calc_size_from_shape<int32_t>(4, shape)); assert_values_match("size", "%d", 210, ndarray_util::calc_size_from_shape<int32_t>(4, shape));
} }
@ -267,9 +335,6 @@ void test_ndslice_1() {
assert dst_ndarray[0, 1] == 7.0 assert dst_ndarray[0, 1] == 7.0
assert dst_ndarray[1, 0] == 9.0 assert dst_ndarray[1, 0] == 9.0
assert dst_ndarray[1, 1] == 11.0 assert dst_ndarray[1, 1] == 11.0
dst_ndarray[1, 0] == 99 # Write to `dst_ndarray`
assert ndarray[1, 3] == 99 # `ndarray` also updates!!
``` ```
*/ */
BEGIN_TEST(); BEGIN_TEST();
@ -410,6 +475,160 @@ void test_ndslice_2() {
assert_values_match("dst_ndarray[1]", "%f", 9.0, *((double *) dst_ndarray.get_pelement((int32_t[dst_ndims]) { 1 }))); assert_values_match("dst_ndarray[1]", "%f", 9.0, *((double *) dst_ndarray.get_pelement((int32_t[dst_ndims]) { 1 })));
} }
void test_can_broadcast_shape() {
BEGIN_TEST();
assert_values_match(
"can_broadcast_shape_to([3], [1, 1, 1, 1, 3]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 3 }, 5, (int32_t[]) { 1, 1, 1, 1, 3 })
);
assert_values_match(
"can_broadcast_shape_to([3], [3, 1]) == false",
"%d",
false,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 3 }, 2, (int32_t[]) { 3, 1 }));
assert_values_match(
"can_broadcast_shape_to([3], [3]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 3 }, 1, (int32_t[]) { 3 }));
assert_values_match(
"can_broadcast_shape_to([1], [3]) == false",
"%d",
false,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 1 }, 1, (int32_t[]) { 3 }));
assert_values_match(
"can_broadcast_shape_to([1], [1]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 1 }, 1, (int32_t[]) { 1 }));
assert_values_match(
"can_broadcast_shape_to([256, 256, 3], [256, 1, 3]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(3, (int32_t[]) { 256, 256, 3 }, 3, (int32_t[]) { 256, 1, 3 })
);
assert_values_match(
"can_broadcast_shape_to([256, 256, 3], [3]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(3, (int32_t[]) { 256, 256, 3 }, 1, (int32_t[]) { 3 })
);
assert_values_match(
"can_broadcast_shape_to([256, 256, 3], [2]) == false",
"%d",
false,
ndarray_util::can_broadcast_shape_to(3, (int32_t[]) { 256, 256, 3 }, 1, (int32_t[]) { 2 })
);
assert_values_match(
"can_broadcast_shape_to([256, 256, 3], [1]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(3, (int32_t[]) { 256, 256, 3 }, 1, (int32_t[]) { 1 })
);
// In cases when the shapes contain zero(es)
assert_values_match(
"can_broadcast_shape_to([0], [1]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 0 }, 1, (int32_t[]) { 1 })
);
assert_values_match(
"can_broadcast_shape_to([0], [2]) == false",
"%d",
false,
ndarray_util::can_broadcast_shape_to(1, (int32_t[]) { 0 }, 1, (int32_t[]) { 2 })
);
assert_values_match(
"can_broadcast_shape_to([0, 4, 0, 0], [1]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(4, (int32_t[]) { 0, 4, 0, 0 }, 1, (int32_t[]) { 1 })
);
assert_values_match(
"can_broadcast_shape_to([0, 4, 0, 0], [1, 1, 1, 1]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(4, (int32_t[]) { 0, 4, 0, 0 }, 4, (int32_t[]) { 1, 1, 1, 1 })
);
assert_values_match(
"can_broadcast_shape_to([0, 4, 0, 0], [1, 4, 1, 1]) == true",
"%d",
true,
ndarray_util::can_broadcast_shape_to(4, (int32_t[]) { 0, 4, 0, 0 }, 4, (int32_t[]) { 1, 4, 1, 1 })
);
assert_values_match(
"can_broadcast_shape_to([4, 3], [0, 3]) == false",
"%d",
false,
ndarray_util::can_broadcast_shape_to(2, (int32_t[]) { 4, 3 }, 2, (int32_t[]) { 0, 3 })
);
assert_values_match(
"can_broadcast_shape_to([4, 3], [0, 0]) == false",
"%d",
false,
ndarray_util::can_broadcast_shape_to(2, (int32_t[]) { 4, 3 }, 2, (int32_t[]) { 0, 0 })
);
}
void test_ndarray_broadcast_1() {
/*
# array = np.array([[19.9, 29.9, 39.9, 49.9]], dtype=np.float64)
# >>> [[19.9 29.9 39.9 49.9]]
#
# array = np.broadcast_to(array, (2, 3, 4))
# >>> [[[19.9 29.9 39.9 49.9]
# >>> [19.9 29.9 39.9 49.9]
# >>> [19.9 29.9 39.9 49.9]]
# >>> [[19.9 29.9 39.9 49.9]
# >>> [19.9 29.9 39.9 49.9]
# >>> [19.9 29.9 39.9 49.9]]]
#
# assery array.strides == (0, 0, 8)
*/
BEGIN_TEST();
double in_data[4] = { 19.9, 29.9, 39.9, 49.9 };
const int32_t in_ndims = 2;
int32_t in_shape[in_ndims] = {1, 4};
int32_t in_strides[in_ndims] = {};
NDArray<int32_t> ndarray = {
.data = (uint8_t*) in_data,
.itemsize = sizeof(double),
.ndims = in_ndims,
.shape = in_shape,
.strides = in_strides
};
ndarray.set_strides_by_shape();
const int32_t dst_ndims = 3;
int32_t dst_shape[dst_ndims] = {2, 3, 4};
int32_t dst_strides[dst_ndims] = {};
NDArray<int32_t> dst_ndarray = {
.ndims = dst_ndims,
.shape = dst_shape,
.strides = dst_strides
};
ndarray.broadcast_to(&dst_ndarray);
assert_arrays_match("dst_ndarray->strides", "%d", dst_ndims, (int32_t[]) { 0, 0, 8 }, dst_ndarray.strides);
assert_values_match("dst_ndarray[0, 0, 0]", "%f", 19.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 0, 0})));
assert_values_match("dst_ndarray[0, 0, 1]", "%f", 29.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 0, 1})));
assert_values_match("dst_ndarray[0, 0, 2]", "%f", 39.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 0, 2})));
assert_values_match("dst_ndarray[0, 0, 3]", "%f", 49.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 0, 3})));
assert_values_match("dst_ndarray[0, 1, 0]", "%f", 19.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 1, 0})));
assert_values_match("dst_ndarray[0, 1, 1]", "%f", 29.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 1, 1})));
assert_values_match("dst_ndarray[0, 1, 2]", "%f", 39.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 1, 2})));
assert_values_match("dst_ndarray[0, 1, 3]", "%f", 49.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {0, 1, 3})));
assert_values_match("dst_ndarray[1, 2, 3]", "%f", 49.9, *((double*) dst_ndarray.get_pelement((int32_t[]) {1, 2, 3})));
}
int main() { int main() {
test_calc_size_from_shape_normal(); test_calc_size_from_shape_normal();
test_calc_size_from_shape_has_zero(); test_calc_size_from_shape_has_zero();
@ -423,5 +642,7 @@ int main() {
test_slice_4(); test_slice_4();
test_ndslice_1(); test_ndslice_1();
test_ndslice_2(); test_ndslice_2();
test_can_broadcast_shape();
test_ndarray_broadcast_1();
return 0; return 0;
} }

View File

@ -30,6 +30,7 @@ namespace {
*death = 0; // TODO: address 0 on hardware might be writable? *death = 0; // TODO: address 0 on hardware might be writable?
} }
// TODO: Make this a macro and allow it to be toggled on/off (e.g., debug vs release)
void irrt_assert(bool condition) { void irrt_assert(bool condition) {
if (!condition) irrt_panic(); if (!condition) irrt_panic();
} }