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

WIP

This commit is contained in:
lyken 2024-07-10 00:38:01 +08:00
parent 4209ad0dff
commit 43e9a9539d
18 changed files with 1263 additions and 989 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
compile_irrt.sh Executable file
View File

@ -0,0 +1,3 @@
#!/usr/bin/env bash
clang-irrt --target=wasm32 -x c++ -fno-discard-value-names -fno-exceptions -fno-rtti -O0 -emit-llvm -S -Wall -Wextra nac3core/irrt/irrt.cpp
clang -x c++ -fno-discard-value-names -fno-exceptions -fno-rtti -O0 -emit-llvm -S -Wall -Wextra nac3core/irrt/irrt_test.cpp

5
flake.nix
View File

@ -41,7 +41,7 @@
'';
installPhase =
''
PYTHON_SITEPACKAGES=$out/${pkgs.python3Packages.python.sitePackages}
u PYTHON_SITEPACKAGES=$out/${pkgs.python3Packages.python.sitePackages}
mkdir -p $PYTHON_SITEPACKAGES
cp target/x86_64-unknown-linux-gnu/release/libnac3artiq.so $PYTHON_SITEPACKAGES/nac3artiq.so
@ -163,7 +163,10 @@
clippy
pre-commit
rustfmt
rust-analyzer
];
# https://nixos.wiki/wiki/Rust#Shell.nix_example
RUST_SRC_PATH = "${pkgs.rust.packages.stable.rustPlatform.rustLibSrc}";
};
devShells.x86_64-linux.msys2 = pkgs.mkShell {
name = "nac3-dev-shell-msys2";

2
nac3core/build.rs
View File

@ -31,6 +31,7 @@ fn compile_irrt(irrt_dir: &Path, out_dir: &Path) {
"-S",
"-Wall",
"-Wextra",
"-Werror=return-type",
"-I",
irrt_dir.to_str().unwrap(),
"-o",
@ -100,6 +101,7 @@ fn compile_irrt_test(irrt_dir: &Path, out_dir: &Path) {
"-O0",
"-Wall",
"-Wextra",
"-Werror=return-type",
"-lm", // for `tgamma()`, `lgamma()`
"-o",
exe_path.to_str().unwrap(),

6
nac3core/irrt/irrt.cpp
View File

@ -1,3 +1,5 @@
#include "irrt.hpp"
#include "irrt_everything.hpp"
// All the implementations are from `irrt.hpp`
/*
This file will be read by `clang-irrt` to conveniently produce LLVM IR for `nac3core/codegen`.
*/

244
nac3core/irrt/irrt.hpp → nac3core/irrt/irrt_basic.hpp
View File

@ -1,28 +1,19 @@
#ifndef IRRT_DONT_TYPEDEF_INTS
typedef _BitInt(8) int8_t;
typedef unsigned _BitInt(8) uint8_t;
typedef _BitInt(32) int32_t;
typedef unsigned _BitInt(32) uint32_t;
typedef _BitInt(64) int64_t;
typedef unsigned _BitInt(64) uint64_t;
#endif
#pragma once
#include "irrt_utils.hpp"
#include "irrt_typedefs.hpp"
/*
This header contains IRRT implementations
that do not deserved to be categorized (e.g., into numpy, etc.)
Check out other *.hpp files before including them here!!
*/
// NDArray indices are always `uint32_t`.
typedef uint32_t NDIndex;
// The type of an index or a value describing the length of a range/slice is
// always `int32_t`.
typedef int32_t SliceIndex;
template <typename T>
static T max(T a, T b) {
return a > b ? a : b;
}
template <typename T>
static T min(T a, T b) {
return a > b ? b : a;
}
// adapted from GNU Scientific Library: https://git.savannah.gnu.org/cgit/gsl.git/tree/sys/pow_int.c
// need to make sure `exp >= 0` before calling this function
template <typename T>
@ -39,119 +30,6 @@ static T __nac3_int_exp_impl(T base, T exp) {
return res;
}
template <typename SizeT>
static SizeT __nac3_ndarray_calc_size_impl(
const SizeT *list_data,
SizeT list_len,
SizeT begin_idx,
SizeT end_idx
) {
__builtin_assume(end_idx <= list_len);
SizeT num_elems = 1;
for (SizeT i = begin_idx; i < end_idx; ++i) {
SizeT val = list_data[i];
__builtin_assume(val > 0);
num_elems *= val;
}
return num_elems;
}
template <typename SizeT>
static void __nac3_ndarray_calc_nd_indices_impl(
SizeT index,
const SizeT *dims,
SizeT num_dims,
NDIndex *idxs
) {
SizeT stride = 1;
for (SizeT dim = 0; dim < num_dims; dim++) {
SizeT i = num_dims - dim - 1;
__builtin_assume(dims[i] > 0);
idxs[i] = (index / stride) % dims[i];
stride *= dims[i];
}
}
template <typename SizeT>
static SizeT __nac3_ndarray_flatten_index_impl(
const SizeT *dims,
SizeT num_dims,
const NDIndex *indices,
SizeT num_indices
) {
SizeT idx = 0;
SizeT stride = 1;
for (SizeT i = 0; i < num_dims; ++i) {
SizeT ri = num_dims - i - 1;
if (ri < num_indices) {
idx += stride * indices[ri];
}
__builtin_assume(dims[i] > 0);
stride *= dims[ri];
}
return idx;
}
template <typename SizeT>
static void __nac3_ndarray_calc_broadcast_impl(
const SizeT *lhs_dims,
SizeT lhs_ndims,
const SizeT *rhs_dims,
SizeT rhs_ndims,
SizeT *out_dims
) {
SizeT max_ndims = lhs_ndims > rhs_ndims ? lhs_ndims : rhs_ndims;
for (SizeT i = 0; i < max_ndims; ++i) {
const SizeT *lhs_dim_sz = i < lhs_ndims ? &lhs_dims[lhs_ndims - i - 1] : nullptr;
const SizeT *rhs_dim_sz = i < rhs_ndims ? &rhs_dims[rhs_ndims - i - 1] : nullptr;
SizeT *out_dim = &out_dims[max_ndims - i - 1];
if (lhs_dim_sz == nullptr) {
*out_dim = *rhs_dim_sz;
} else if (rhs_dim_sz == nullptr) {
*out_dim = *lhs_dim_sz;
} else if (*lhs_dim_sz == 1) {
*out_dim = *rhs_dim_sz;
} else if (*rhs_dim_sz == 1) {
*out_dim = *lhs_dim_sz;
} else if (*lhs_dim_sz == *rhs_dim_sz) {
*out_dim = *lhs_dim_sz;
} else {
__builtin_unreachable();
}
}
}
template <typename SizeT>
static void __nac3_ndarray_calc_broadcast_idx_impl(
const SizeT *src_dims,
SizeT src_ndims,
const NDIndex *in_idx,
NDIndex *out_idx
) {
for (SizeT i = 0; i < src_ndims; ++i) {
SizeT src_i = src_ndims - i - 1;
out_idx[src_i] = src_dims[src_i] == 1 ? 0 : in_idx[src_i];
}
}
template<typename SizeT>
static void __nac3_ndarray_strides_from_shape_impl(
SizeT ndims,
SizeT *shape,
SizeT *dst_strides
) {
SizeT stride_product = 1;
for (SizeT i = 0; i < ndims; i++) {
int dim_i = ndims - i - 1;
dst_strides[dim_i] = stride_product;
stride_product *= shape[dim_i];
}
}
extern "C" {
#define DEF_nac3_int_exp_(T) \
T __nac3_int_exp_##T(T base, T exp) {\
@ -334,104 +212,4 @@ extern "C" {
return j0(x);
}
uint32_t __nac3_ndarray_calc_size(
const uint32_t *list_data,
uint32_t list_len,
uint32_t begin_idx,
uint32_t end_idx
) {
return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
}
uint64_t __nac3_ndarray_calc_size64(
const uint64_t *list_data,
uint64_t list_len,
uint64_t begin_idx,
uint64_t end_idx
) {
return __nac3_ndarray_calc_size_impl(list_data, list_len, begin_idx, end_idx);
}
void __nac3_ndarray_calc_nd_indices(
uint32_t index,
const uint32_t* dims,
uint32_t num_dims,
NDIndex* idxs
) {
__nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
}
void __nac3_ndarray_calc_nd_indices64(
uint64_t index,
const uint64_t* dims,
uint64_t num_dims,
NDIndex* idxs
) {
__nac3_ndarray_calc_nd_indices_impl(index, dims, num_dims, idxs);
}
uint32_t __nac3_ndarray_flatten_index(
const uint32_t* dims,
uint32_t num_dims,
const NDIndex* indices,
uint32_t num_indices
) {
return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
}
uint64_t __nac3_ndarray_flatten_index64(
const uint64_t* dims,
uint64_t num_dims,
const NDIndex* indices,
uint64_t num_indices
) {
return __nac3_ndarray_flatten_index_impl(dims, num_dims, indices, num_indices);
}
void __nac3_ndarray_calc_broadcast(
const uint32_t *lhs_dims,
uint32_t lhs_ndims,
const uint32_t *rhs_dims,
uint32_t rhs_ndims,
uint32_t *out_dims
) {
return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
}
void __nac3_ndarray_calc_broadcast64(
const uint64_t *lhs_dims,
uint64_t lhs_ndims,
const uint64_t *rhs_dims,
uint64_t rhs_ndims,
uint64_t *out_dims
) {
return __nac3_ndarray_calc_broadcast_impl(lhs_dims, lhs_ndims, rhs_dims, rhs_ndims, out_dims);
}
void __nac3_ndarray_calc_broadcast_idx(
const uint32_t *src_dims,
uint32_t src_ndims,
const NDIndex *in_idx,
NDIndex *out_idx
) {
__nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
}
void __nac3_ndarray_calc_broadcast_idx64(
const uint64_t *src_dims,
uint64_t src_ndims,
const NDIndex *in_idx,
NDIndex *out_idx
) {
__nac3_ndarray_calc_broadcast_idx_impl(src_dims, src_ndims, in_idx, out_idx);
}
void __nac3_ndarray_strides_from_shape(uint32_t ndims, uint32_t* shape, uint32_t* dst_strides) {
__nac3_ndarray_strides_from_shape_impl(ndims, shape, dst_strides);
}
void __nac3_ndarray_strides_from_shape64(uint64_t ndims, uint64_t* shape, uint64_t* dst_strides) {
__nac3_ndarray_strides_from_shape_impl(ndims, shape, dst_strides);
}
}

11
nac3core/irrt/irrt_everything.hpp Normal file
View File

@ -0,0 +1,11 @@
#pragma once
#include "irrt_basic.hpp"
#include "irrt_numpy_ndarray.hpp"
/*
All IRRT implementations.
We don't have any pre-compiled objects, so we are writing all implementations in headers and
concatenate them with `#include` into one massive source file that contains all the IRRT stuff.
*/

196
nac3core/irrt/irrt_numpy_ndarray.hpp Normal file
View File

@ -0,0 +1,196 @@
#pragma once
#include "irrt_utils.hpp"
#include "irrt_typedefs.hpp"
/*
NDArray-related implementations.
`*/
// NDArray indices are always `uint32_t`.
using NDIndex = uint32_t;
namespace {
namespace ndarray_util {
// 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 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;
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>
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_value_at_pelement(uint8_t* pelement, uint8_t* pvalue) {
__builtin_memcpy(pelement, pvalue, itemsize);
}
uint8_t* get_pelement(SizeT *indices) {
uint8_t* element = data;
for (SizeT dim_i = 0; dim_i < ndims; dim_i++)
element += indices[dim_i] * strides[dim_i] * itemsize;
return element;
}
// Is the given `indices` valid/in-bounds?
bool in_bounds(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(uint8_t* pvalue) {
NDArrayIndicesIter<SizeT> iter;
iter.ndims = this->ndims;
iter.shape = this->shape;
iter.indices = (SizeT*) __builtin_alloca(sizeof(SizeT) * ndims);
iter.set_indices_zero();
for (SizeT i = 0; i < this->size(); i++, iter.next()) {
uint8_t* pelement = get_pelement(iter.indices);
set_value_at_pelement(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(ndims, strides, shape);
}
// https://numpy.org/doc/stable/reference/generated/numpy.eye.html
void set_to_eye(SizeT k, uint8_t* zero_pvalue, 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(indices);
set_value_at_pelement(pelement, one_pvalue);
}
}
};
}
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_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);
}
}

171
nac3core/irrt/irrt_test.cpp
View File

@ -2,17 +2,21 @@
#include <cstdio>
#include <cstdlib>
// set `IRRT_DONT_TYPEDEF_INTS` because `cstdint` has it all
#define IRRT_DONT_TYPEDEF_INTS
#include "irrt.hpp"
#include "irrt_everything.hpp"
static void __test_fail(const char *file, int line) {
// NOTE: Try to make the location info follow a format that
// VSCode/other IDEs would recognize as a clickable URL.
printf("[!] test_fail() invoked at %s:%d", file, line);
namespace {
static void test_fail() {
printf("[!] Test failed\n");
exit(1);
}
#define test_fail() __test_fail(__FILE__, __LINE__);
static void __begin_test(const char* function_name, const char* file, int line) {
printf("######### Running %s @ %s:%d\n", function_name, file, line);
}
#define BEGIN_TEST() __begin_test(__FUNCTION__, __FILE__, __LINE__)
template <typename T>
bool arrays_match(int len, T *as, T *bs) {
@ -23,40 +27,163 @@ bool arrays_match(int len, T *as, T *bs) {
}
template <typename T>
void debug_print_array(const char* format, int len, T *as) {
void debug_print_array(const char* format, int len, T* as) {
printf("[");
for (int i = 0; i < len; i++) {
if (i != 0) printf(", ");
printf(format, as[i]);
}
printf("]\n");
printf("]");
}
template <typename T>
bool assert_arrays_match(const char *label, const char *format, int len, T *expected, T *got) {
auto match = arrays_match(len, expected, got);
if (!match) {
void assert_arrays_match(const char* label, const char* format, int len, T* expected, T* got) {
if (!arrays_match(len, expected, got)) {
printf("expected %s: ", label);
debug_print_array(format, len, expected);
printf("\n");
printf("got %s: ", label);
debug_print_array(format, len, got);
printf("\n");
test_fail();
}
return match;
}
static void test_strides_from_shape() {
const uint64_t ndims = 4;
uint64_t shape[ndims] = { 999, 3, 5, 7 };
uint64_t strides[ndims] = { 0 };
__nac3_ndarray_strides_from_shape64(ndims, shape, strides);
template <typename T>
void assert_values_match(const char* label, const char* format, T expected, T got) {
if (expected != got) {
printf("expected %s: ", label);
printf(format, expected);
printf("\n");
printf("got %s: ", label);
printf(format, got);
printf("\n");
test_fail();
}
}
uint64_t expected_strides[ndims] = { 3*5*7, 5*7, 7, 1 };
if (!assert_arrays_match("strides", "%u", ndims, expected_strides, strides)) test_fail();
void test_calc_size_from_shape_normal() {
// Test shapes with normal values
BEGIN_TEST();
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));
}
void test_calc_size_from_shape_has_zero() {
// Test shapes with 0 in them
BEGIN_TEST();
int32_t shape[4] = { 2, 0, 5, 7 };
assert_values_match("size", "%d", 0, ndarray_util::calc_size_from_shape<int32_t>(4, shape));
}
void test_set_strides_by_shape() {
// Test `set_strides_by_shape()`
BEGIN_TEST();
int32_t shape[4] = { 99, 3, 5, 7 };
int32_t strides[4] = { 0 };
ndarray_util::set_strides_by_shape(4, strides, shape);
int32_t expected_strides[4] = { 105, 35, 7, 1 };
assert_arrays_match("strides", "%u", 4u, expected_strides, strides);
}
void test_ndarray_indices_iter_normal() {
// Test NDArrayIndicesIter normal behavior
BEGIN_TEST();
int32_t shape[3] = { 1, 2, 3 };
int32_t indices[3] = { 0, 0, 0 };
auto iter = NDArrayIndicesIter<int32_t> {
.ndims = 3u,
.shape = shape,
.indices = indices
};
assert_arrays_match("indices #0", "%u", 3u, iter.indices, (int32_t[3]) { 0, 0, 0 });
iter.next();
assert_arrays_match("indices #1", "%u", 3u, iter.indices, (int32_t[3]) { 0, 0, 1 });
iter.next();
assert_arrays_match("indices #2", "%u", 3u, iter.indices, (int32_t[3]) { 0, 0, 2 });
iter.next();
assert_arrays_match("indices #3", "%u", 3u, iter.indices, (int32_t[3]) { 0, 1, 0 });
iter.next();
assert_arrays_match("indices #4", "%u", 3u, iter.indices, (int32_t[3]) { 0, 1, 1 });
iter.next();
assert_arrays_match("indices #5", "%u", 3u, iter.indices, (int32_t[3]) { 0, 1, 2 });
iter.next();
assert_arrays_match("indices #6", "%u", 3u, iter.indices, (int32_t[3]) { 0, 0, 0 }); // Loops back
iter.next();
assert_arrays_match("indices #7", "%u", 3u, iter.indices, (int32_t[3]) { 0, 0, 1 });
}
void test_ndarray_fill_generic() {
// Test ndarray fill_generic
BEGIN_TEST();
// Choose a type that's neither int32_t nor uint64_t (candidates of SizeT) to spice it up
// Also make all the octets non-zero, to see if `memcpy` in `fill_generic` is working perfectly.
uint16_t fill_value = 0xFACE;
uint16_t in_data[6] = { 100, 101, 102, 103, 104, 105 }; // Fill `data` with values that != `999`
int32_t in_itemsize = sizeof(uint16_t);
const int32_t in_ndims = 2;
int32_t in_shape[in_ndims] = { 2, 3 };
int32_t in_strides[in_ndims] = {};
NDArray<int32_t> ndarray = {
.data = (uint8_t*) in_data,
.itemsize = in_itemsize,
.ndims = in_ndims,
.shape = in_shape,
.strides = in_strides,
};
ndarray.set_strides_by_shape();
ndarray.fill_generic((uint8_t*) &fill_value); // `fill_generic` here
uint16_t expected_data[6] = { fill_value, fill_value, fill_value, fill_value, fill_value, fill_value };
assert_arrays_match("data", "0x%hX", 6, expected_data, in_data);
}
void test_ndarray_set_to_eye() {
double in_data[9] = { 99.0, 99.0, 99.0, 99.0, 99.0, 99.0, 99.0, 99.0, 99.0 };
int32_t in_itemsize = sizeof(double);
const int32_t in_ndims = 2;
int32_t in_shape[in_ndims] = { 3, 3 };
int32_t in_strides[in_ndims] = {};
NDArray<int32_t> ndarray = {
.data = (uint8_t*) in_data,
.itemsize = in_itemsize,
.ndims = in_ndims,
.shape = in_shape,
.strides = in_strides,
};
ndarray.set_strides_by_shape();
double zero = 0.0;
double one = 1.0;
ndarray.set_to_eye(1, (uint8_t*) &zero, (uint8_t*) &one);
assert_values_match("in_data[0]", "%f", 0.0, in_data[0]);
assert_values_match("in_data[1]", "%f", 1.0, in_data[1]);
assert_values_match("in_data[2]", "%f", 0.0, in_data[2]);
assert_values_match("in_data[3]", "%f", 0.0, in_data[3]);
assert_values_match("in_data[4]", "%f", 0.0, in_data[4]);
assert_values_match("in_data[5]", "%f", 1.0, in_data[5]);
assert_values_match("in_data[6]", "%f", 0.0, in_data[6]);
assert_values_match("in_data[7]", "%f", 0.0, in_data[7]);
assert_values_match("in_data[8]", "%f", 0.0, in_data[8]);
}
}
int main() {
test_strides_from_shape();
test_calc_size_from_shape_normal();
test_calc_size_from_shape_has_zero();
test_set_strides_by_shape();
test_ndarray_indices_iter_normal();
test_ndarray_fill_generic();
test_ndarray_set_to_eye();
return 0;
}

12
nac3core/irrt/irrt_typedefs.hpp Normal file
View File

@ -0,0 +1,12 @@
#pragma once
// This is made toggleable since `irrt_test.cpp` itself would include
// headers that define the `int_t` family.
#ifndef IRRT_DONT_TYPEDEF_INTS
typedef _BitInt(8) int8_t;
typedef unsigned _BitInt(8) uint8_t;
typedef _BitInt(32) int32_t;
typedef unsigned _BitInt(32) uint32_t;
typedef _BitInt(64) int64_t;
typedef unsigned _BitInt(64) uint64_t;
#endif

11
nac3core/irrt/irrt_utils.hpp Normal file
View File

@ -0,0 +1,11 @@
#pragma once
template <typename T>
static T max(T a, T b) {
return a > b ? a : b;
}
template <typename T>
static T min(T a, T b) {
return a > b ? b : a;
}

162
nac3core/src/codegen/builtin_fns.rs
View File

@ -702,53 +702,54 @@ pub fn call_numpy_min<'ctx, G: CodeGenerator + ?Sized>(
BasicValueEnum::PointerValue(n)
if a_ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id()) =>
{
let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, a_ty);
let llvm_ndarray_ty = ctx.get_llvm_type(generator, elem_ty);
todo!()
// let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, a_ty);
// let llvm_ndarray_ty = ctx.get_llvm_type(generator, elem_ty);
let n = NDArrayValue::from_ptr_val(n, llvm_usize, None);
let n_sz = irrt::call_ndarray_calc_size(generator, ctx, &n.dim_sizes(), (None, None));
if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
let n_sz_eqz = ctx
.builder
.build_int_compare(IntPredicate::NE, n_sz, n_sz.get_type().const_zero(), "")
.unwrap();
// let n = NDArrayValue::from_ptr_val(n, llvm_usize, None);
// let n_sz = irrt::call_ndarray_calc_size(generator, ctx, &n.dim_sizes(), (None, None));
// if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
// let n_sz_eqz = ctx
// .builder
// .build_int_compare(IntPredicate::NE, n_sz, n_sz.get_type().const_zero(), "")
// .unwrap();
ctx.make_assert(
generator,
n_sz_eqz,
"0:ValueError",
"zero-size array to reduction operation minimum which has no identity",
[None, None, None],
ctx.current_loc,
);
}
// ctx.make_assert(
// generator,
// n_sz_eqz,
// "0:ValueError",
// "zero-size array to reduction operation minimum which has no identity",
// [None, None, None],
// ctx.current_loc,
// );
// }
let accumulator_addr = generator.gen_var_alloc(ctx, llvm_ndarray_ty, None)?;
unsafe {
let identity =
n.data().get_unchecked(ctx, generator, &llvm_usize.const_zero(), None);
ctx.builder.build_store(accumulator_addr, identity).unwrap();
}
// let accumulator_addr = generator.gen_var_alloc(ctx, llvm_ndarray_ty, None)?;
// unsafe {
// let identity =
// n.data().get_unchecked(ctx, generator, &llvm_usize.const_zero(), None);
// ctx.builder.build_store(accumulator_addr, identity).unwrap();
// }
gen_for_callback_incrementing(
generator,
ctx,
llvm_usize.const_int(1, false),
(n_sz, false),
|generator, ctx, _, idx| {
let elem = unsafe { n.data().get_unchecked(ctx, generator, &idx, None) };
// gen_for_callback_incrementing(
// generator,
// ctx,
// llvm_usize.const_int(1, false),
// (n_sz, false),
// |generator, ctx, _, idx| {
// let elem = unsafe { n.data().get_unchecked(ctx, generator, &idx, None) };
let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
let result = call_min(ctx, (elem_ty, accumulator), (elem_ty, elem));
ctx.builder.build_store(accumulator_addr, result).unwrap();
// let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
// let result = call_min(ctx, (elem_ty, accumulator), (elem_ty, elem));
// ctx.builder.build_store(accumulator_addr, result).unwrap();
Ok(())
},
llvm_usize.const_int(1, false),
)?;
// Ok(())
// },
// llvm_usize.const_int(1, false),
// )?;
let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
accumulator
// let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
// accumulator
}
_ => unsupported_type(ctx, FN_NAME, &[a_ty]),
@ -920,53 +921,54 @@ pub fn call_numpy_max<'ctx, G: CodeGenerator + ?Sized>(
BasicValueEnum::PointerValue(n)
if a_ty.obj_id(&ctx.unifier).is_some_and(|id| id == PrimDef::NDArray.id()) =>
{
let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, a_ty);
let llvm_ndarray_ty = ctx.get_llvm_type(generator, elem_ty);
todo!()
// let (elem_ty, _) = unpack_ndarray_var_tys(&mut ctx.unifier, a_ty);
// let llvm_ndarray_ty = ctx.get_llvm_type(generator, elem_ty);
let n = NDArrayValue::from_ptr_val(n, llvm_usize, None);
let n_sz = irrt::call_ndarray_calc_size(generator, ctx, &n.dim_sizes(), (None, None));
if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
let n_sz_eqz = ctx
.builder
.build_int_compare(IntPredicate::NE, n_sz, n_sz.get_type().const_zero(), "")
.unwrap();
// let n = NDArrayValue::from_ptr_val(n, llvm_usize, None);
// let n_sz = irrt::call_ndarray_calc_size(generator, ctx, &n.dim_sizes(), (None, None));
// if ctx.registry.llvm_options.opt_level == OptimizationLevel::None {
// let n_sz_eqz = ctx
// .builder
// .build_int_compare(IntPredicate::NE, n_sz, n_sz.get_type().const_zero(), "")
// .unwrap();
ctx.make_assert(
generator,
n_sz_eqz,
"0:ValueError",
"zero-size array to reduction operation minimum which has no identity",
[None, None, None],
ctx.current_loc,
);
}
// ctx.make_assert(
// generator,
// n_sz_eqz,
// "0:ValueError",
// "zero-size array to reduction operation minimum which has no identity",
// [None, None, None],
// ctx.current_loc,
// );
// }
let accumulator_addr = generator.gen_var_alloc(ctx, llvm_ndarray_ty, None)?;
unsafe {
let identity =
n.data().get_unchecked(ctx, generator, &llvm_usize.const_zero(), None);
ctx.builder.build_store(accumulator_addr, identity).unwrap();
}
// let accumulator_addr = generator.gen_var_alloc(ctx, llvm_ndarray_ty, None)?;
// unsafe {
// let identity =
// n.data().get_unchecked(ctx, generator, &llvm_usize.const_zero(), None);
// ctx.builder.build_store(accumulator_addr, identity).unwrap();
// }
gen_for_callback_incrementing(
generator,
ctx,
llvm_usize.const_int(1, false),
(n_sz, false),
|generator, ctx, _, idx| {
let elem = unsafe { n.data().get_unchecked(ctx, generator, &idx, None) };
// gen_for_callback_incrementing(
// generator,
// ctx,
// llvm_usize.const_int(1, false),
// (n_sz, false),
// |generator, ctx, _, idx| {
// let elem = unsafe { n.data().get_unchecked(ctx, generator, &idx, None) };
let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
let result = call_max(ctx, (elem_ty, accumulator), (elem_ty, elem));
ctx.builder.build_store(accumulator_addr, result).unwrap();
// let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
// let result = call_max(ctx, (elem_ty, accumulator), (elem_ty, elem));
// ctx.builder.build_store(accumulator_addr, result).unwrap();
Ok(())
},
llvm_usize.const_int(1, false),
)?;
// Ok(())
// },
// llvm_usize.const_int(1, false),
// )?;
let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
accumulator
// let accumulator = ctx.builder.build_load(accumulator_addr, "").unwrap();
// accumulator
}
_ => unsupported_type(ctx, FN_NAME, &[a_ty]),

336
nac3core/src/codegen/classes.rs
View File

@ -1,8 +1,6 @@
use crate::codegen::{
irrt::{call_ndarray_calc_size, call_ndarray_flatten_index},
llvm_intrinsics::call_int_umin,
stmt::gen_for_callback_incrementing,
CodeGenContext, CodeGenerator,
llvm_intrinsics::call_int_umin, stmt::gen_for_callback_incrementing, CodeGenContext,
CodeGenerator,
};
use inkwell::context::Context;
use inkwell::types::{ArrayType, BasicType, StructType};
@ -12,6 +10,7 @@ use inkwell::{
values::{BasicValueEnum, IntValue, PointerValue},
AddressSpace, IntPredicate,
};
use itertools::Itertools;
/// A LLVM type that is used to represent a non-primitive type in NAC3.
pub trait ProxyType<'ctx>: Into<Self::Base> {
@ -1601,7 +1600,8 @@ impl<'ctx> ArrayLikeValue<'ctx> for NDArrayDataProxy<'ctx, '_> {
ctx: &CodeGenContext<'ctx, '_>,
generator: &G,
) -> IntValue<'ctx> {
call_ndarray_calc_size(generator, ctx, &self.as_slice_value(ctx, generator), (None, None))
todo!()
// call_ndarray_calc_size(generator, ctx, &self.as_slice_value(ctx, generator), (None, None))
}
}
@ -1675,17 +1675,19 @@ impl<'ctx, Index: UntypedArrayLikeAccessor<'ctx>> ArrayLikeIndexer<'ctx, Index>
indices_elem_ty.get_bit_width()
);
let index = call_ndarray_flatten_index(generator, ctx, *self.0, indices);
todo!()
unsafe {
ctx.builder
.build_in_bounds_gep(
self.base_ptr(ctx, generator),
&[index],
name.unwrap_or_default(),
)
.unwrap()
}
// let index = call_ndarray_flatten_index(generator, ctx, *self.0, indices);
// unsafe {
// ctx.builder
// .build_in_bounds_gep(
// self.base_ptr(ctx, generator),
// &[index],
// name.unwrap_or_default(),
// )
// .unwrap()
// }
}
fn ptr_offset<G: CodeGenerator + ?Sized>(
@ -1761,3 +1763,307 @@ impl<'ctx, Index: UntypedArrayLikeAccessor<'ctx>> UntypedArrayLikeMutator<'ctx,
for NDArrayDataProxy<'ctx, '_>
{
}
#[derive(Debug, Clone, Copy)]
pub struct StructField<'ctx> {
/// The GEP index of this struct field.
pub gep_index: u32,
/// Name of this struct field.
///
/// Used for generating names.
pub name: &'static str,
/// The type of this struct field.
pub ty: BasicTypeEnum<'ctx>,
}
pub struct StructFields<'ctx> {
/// Name of the struct.
///
/// Used for generating names.
pub name: &'static str,
/// All the [`StructField`]s of this struct.
///
/// **NOTE:** The index position of a [`StructField`]
/// matches the element's [`StructField::index`].
pub fields: Vec<StructField<'ctx>>,
}
struct StructFieldsBuilder<'ctx> {
gep_index_counter: u32,
/// Name of the struct to be built.
name: &'static str,
fields: Vec<StructField<'ctx>>,
}
impl<'ctx> StructField<'ctx> {
pub fn gep(
&self,
ctx: &CodeGenContext<'ctx, '_>,
ptr: PointerValue<'ctx>,
) -> PointerValue<'ctx> {
ctx.builder.build_struct_gep(ptr, self.gep_index, self.name).unwrap()
}
pub fn load(
&self,
ctx: &CodeGenContext<'ctx, '_>,
ptr: PointerValue<'ctx>,
) -> BasicValueEnum<'ctx> {
ctx.builder.build_load(self.gep(ctx, ptr), self.name).unwrap()
}
pub fn store<V>(&self, ctx: &CodeGenContext<'ctx, '_>, ptr: PointerValue<'ctx>, value: V)
where
V: BasicValue<'ctx>,
{
ctx.builder.build_store(ptr, value).unwrap();
}
}
type IsInstanceError = String;
type IsInstanceResult = Result<(), IsInstanceError>;
pub fn check_basic_types_match<'ctx, A, B>(expected: A, got: B) -> IsInstanceResult
where
A: BasicType<'ctx>,
B: BasicType<'ctx>,
{
let expected = expected.as_basic_type_enum();
let got = got.as_basic_type_enum();
// Put those logic into here,
// otherwise there is always a fallback reporting on any kind of mismatch
match (expected, got) {
(BasicTypeEnum::IntType(expected), BasicTypeEnum::IntType(got)) => {
if expected.get_bit_width() != got.get_bit_width() {
return Err(format!(
"Expected IntType ({expected}-bit(s)), got IntType ({got}-bit(s))"
));
}
}
(expected, got) => {
if expected != got {
return Err(format!("Expected {expected}, got {got}"));
}
}
}
Ok(())
}
impl<'ctx> StructFields<'ctx> {
pub fn num_fields(&self) -> u32 {
self.fields.len() as u32
}
pub fn as_struct_type(&self, ctx: &'ctx Context) -> StructType<'ctx> {
let llvm_fields = self.fields.iter().map(|field| field.ty).collect_vec();
ctx.struct_type(llvm_fields.as_slice(), false)
}
pub fn is_type(&self, scrutinee: StructType<'ctx>) -> IsInstanceResult {
// Check scrutinee's number of struct fields
if scrutinee.count_fields() != self.num_fields() {
return Err(format!(
"Expected {expected_count} field(s) in `{struct_name}` type, got {got_count}",
struct_name = self.name,
expected_count = self.num_fields(),
got_count = scrutinee.count_fields(),
));
}
// Check the scrutinee's field types
for field in self.fields.iter() {
let expected_field_ty = field.ty;
let got_field_ty = scrutinee.get_field_type_at_index(field.gep_index).unwrap();
if let Err(field_err) = check_basic_types_match(expected_field_ty, got_field_ty) {
return Err(format!(
"Field GEP index {gep_index} does not match the expected type of ({struct_name}::{field_name}): {field_err}",
gep_index = field.gep_index,
struct_name = self.name,
field_name = field.name,
));
}
}
// Done
Ok(())
}
}
impl<'ctx> StructFieldsBuilder<'ctx> {
fn start(name: &'static str) -> Self {
StructFieldsBuilder { gep_index_counter: 0, name, fields: Vec::new() }
}
fn add_field(&mut self, name: &'static str, ty: BasicTypeEnum<'ctx>) -> StructField<'ctx> {
let index = self.gep_index_counter;
self.gep_index_counter += 1;
StructField { gep_index: index, name, ty }
}
fn end(self) -> StructFields<'ctx> {
StructFields { name: self.name, fields: self.fields }
}
}
#[derive(Debug, Clone, Copy)]
pub struct NpArrayType<'ctx> {
pub size_type: IntType<'ctx>,
pub elem_type: BasicTypeEnum<'ctx>,
}
pub struct NpArrayStructFields<'ctx> {
pub whole_struct: StructFields<'ctx>,
pub data: StructField<'ctx>,
pub itemsize: StructField<'ctx>,
pub ndims: StructField<'ctx>,
pub shape: StructField<'ctx>,
pub strides: StructField<'ctx>,
}
impl<'ctx> NpArrayType<'ctx> {
pub fn new_opaque_elem(
ctx: &CodeGenContext<'ctx, '_>,
size_type: IntType<'ctx>,
) -> NpArrayType<'ctx> {
NpArrayType { size_type, elem_type: ctx.ctx.i8_type().as_basic_type_enum() }
}
pub fn struct_type(&self, ctx: &CodeGenContext<'ctx, '_>) -> StructType<'ctx> {
self.fields().whole_struct.as_struct_type(ctx.ctx)
}
pub fn fields(&self) -> NpArrayStructFields<'ctx> {
let mut builder = StructFieldsBuilder::start("NpArray");
let addrspace = AddressSpace::default();
let byte_type = self.size_type.get_context().i8_type();
// Make sure the struct matches PERFECTLY with that defined in `nac3core/irrt`.
let data = builder.add_field("data", byte_type.ptr_type(addrspace).into());
let itemsize = builder.add_field("itemsize", self.size_type.into());
let ndims = builder.add_field("ndims", self.size_type.into());
let shape = builder.add_field("shape", self.size_type.ptr_type(addrspace).into());
let strides = builder.add_field("strides", self.size_type.ptr_type(addrspace).into());
NpArrayStructFields { whole_struct: builder.end(), data, itemsize, ndims, shape, strides }
}
/// Allocate an `ndarray` on stack, with the following notes:
///
/// - `ndarray.ndims` will be initialized to `in_ndims`.
/// - `ndarray.itemsize` will be initialized to the size of `self.elem_type.size_of()`.
/// - `ndarray.shape` and `ndarray.strides` will be allocated on the stack with number of elements being `in_ndims`,
/// all with empty/uninitialized values.
pub fn alloca(
&self,
ctx: &CodeGenContext<'ctx, '_>,
in_ndims: IntValue<'ctx>,
name: &str,
) -> NpArrayValue<'ctx> {
let fields = self.fields();
let ptr =
ctx.builder.build_alloca(fields.whole_struct.as_struct_type(ctx.ctx), name).unwrap();
// Allocate `in_dims` number of `size_type` on the stack for `shape` and `strides`
let allocated_shape =
ctx.builder.build_array_alloca(fields.shape.ty, in_ndims, "allocated_shape").unwrap();
let allocated_strides = ctx
.builder
.build_array_alloca(fields.strides.ty, in_ndims, "allocated_strides")
.unwrap();
let value = NpArrayValue { ty: *self, ptr };
value.store_ndims(ctx, in_ndims);
value.store_itemsize(ctx, self.elem_type.size_of().unwrap());
value.store_shape(ctx, allocated_shape);
value.store_strides(ctx, allocated_strides);
return value;
}
}
#[derive(Debug, Clone, Copy)]
pub struct NpArrayValue<'ctx> {
pub ty: NpArrayType<'ctx>,
pub ptr: PointerValue<'ctx>,
}
impl<'ctx> NpArrayValue<'ctx> {
pub fn load_ndims(&self, ctx: &CodeGenContext<'ctx, '_>) -> IntValue<'ctx> {
let field = self.ty.fields().ndims;
field.load(ctx, self.ptr).into_int_value()
}
pub fn store_ndims(&self, ctx: &CodeGenContext<'ctx, '_>, value: IntValue<'ctx>) {
let field = self.ty.fields().ndims;
field.store(ctx, self.ptr, value);
}
pub fn load_itemsize(&self, ctx: &CodeGenContext<'ctx, '_>) -> IntValue<'ctx> {
let field = self.ty.fields().itemsize;
field.load(ctx, self.ptr).into_int_value()
}
pub fn store_itemsize(&self, ctx: &CodeGenContext<'ctx, '_>, value: IntValue<'ctx>) {
let field = self.ty.fields().itemsize;
field.store(ctx, self.ptr, value);
}
pub fn load_shape(&self, ctx: &CodeGenContext<'ctx, '_>) -> PointerValue<'ctx> {
let field = self.ty.fields().shape;
field.load(ctx, self.ptr).into_pointer_value()
}
pub fn store_shape(&self, ctx: &CodeGenContext<'ctx, '_>, value: PointerValue<'ctx>) {
let field = self.ty.fields().shape;
field.store(ctx, self.ptr, value);
}
pub fn load_strides(&self, ctx: &CodeGenContext<'ctx, '_>) -> PointerValue<'ctx> {
let field = self.ty.fields().strides;
field.load(ctx, self.ptr).into_pointer_value()
}
pub fn store_strides(&self, ctx: &CodeGenContext<'ctx, '_>, value: PointerValue<'ctx>) {
let field = self.ty.fields().strides;
field.store(ctx, self.ptr, value);
}
/// TODO: DOCUMENT ME -- NDIMS WOULD NEVER CHANGE!!!!!
pub fn shape_slice(
&self,
ctx: &CodeGenContext<'ctx, '_>,
) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
let field = self.ty.fields().shape;
field.gep(ctx, self.ptr);
let ndims = self.load_ndims(ctx);
TypedArrayLikeAdapter {
adapted: ArraySliceValue(self.ptr, ndims, Some(field.name)),
downcast_fn: Box::new(|_ctx, x| x.into_int_value()),
upcast_fn: Box::new(|_ctx, x| x.as_basic_value_enum()),
}
}
/// TODO: DOCUMENT ME -- NDIMS WOULD NEVER CHANGE!!!!!
pub fn strides_slice(
&self,
ctx: &CodeGenContext<'ctx, '_>,
) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
let field = self.ty.fields().strides;
field.gep(ctx, self.ptr);
let ndims = self.load_ndims(ctx);
TypedArrayLikeAdapter {
adapted: ArraySliceValue(self.ptr, ndims, Some(field.name)),
downcast_fn: Box::new(|_ctx, x| x.into_int_value()),
upcast_fn: Box::new(|_ctx, x| x.as_basic_value_enum()),
}
}
}

419
nac3core/src/codegen/irrt/mod.rs
View File

@ -1,10 +1,14 @@
use crate::typecheck::typedef::Type;
use crate::{
codegen::classes::{NDArrayType, NpArrayType},
typecheck::typedef::Type,
util::SizeVariant,
};
mod test;
use super::{
classes::{
ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue,
ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue, NpArrayValue,
TypedArrayLikeAdapter, UntypedArrayLikeAccessor,
},
llvm_intrinsics, CodeGenContext, CodeGenerator,
@ -16,8 +20,8 @@ use inkwell::{
context::Context,
memory_buffer::MemoryBuffer,
module::Module,
types::{BasicTypeEnum, IntType},
values::{BasicValueEnum, CallSiteValue, FloatValue, IntValue},
types::{BasicType, BasicTypeEnum, FunctionType, IntType, PointerType},
values::{BasicValueEnum, CallSiteValue, FloatValue, FunctionValue, IntValue, PointerValue},
AddressSpace, IntPredicate,
};
use itertools::Either;
@ -565,367 +569,62 @@ pub fn call_j0<'ctx>(ctx: &CodeGenContext<'ctx, '_>, v: FloatValue<'ctx>) -> Flo
.unwrap()
}
/// Generates a call to `__nac3_ndarray_calc_size`. Returns an [`IntValue`] representing the
/// calculated total size.
///
/// * `dims` - An [`ArrayLikeIndexer`] containing the size of each dimension.
/// * `range` - The dimension index to begin and end (exclusively) calculating the dimensions for,
/// or [`None`] if starting from the first dimension and ending at the last dimension respectively.
pub fn call_ndarray_calc_size<'ctx, G, Dims>(
generator: &G,
fn get_size_variant<'ctx>(ty: IntType<'ctx>) -> SizeVariant {
match ty.get_bit_width() {
32 => SizeVariant::Bits32,
64 => SizeVariant::Bits64,
_ => unreachable!("Unsupported int type bit width {}", ty.get_bit_width()),
}
}
fn get_size_type_dependent_function<'ctx, BuildFuncTypeFn>(
ctx: &CodeGenContext<'ctx, '_>,
dims: &Dims,
(begin, end): (Option<IntValue<'ctx>>, Option<IntValue<'ctx>>),
) -> IntValue<'ctx>
size_type: IntType<'ctx>,
base_name: &str,
build_func_type: BuildFuncTypeFn,
) -> FunctionValue<'ctx>
where
G: CodeGenerator + ?Sized,
Dims: ArrayLikeIndexer<'ctx>,
BuildFuncTypeFn: Fn() -> FunctionType<'ctx>,
{
let llvm_usize = generator.get_size_type(ctx.ctx);
let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
let mut fn_name = base_name.to_owned();
match get_size_variant(size_type) {
SizeVariant::Bits32 => {
// The original fn_name is the correct function name
}
SizeVariant::Bits64 => {
// Append "64" at the end, this is the naming convention for 64-bit
fn_name.push_str("64");
}
}
let ndarray_calc_size_fn_name = match llvm_usize.get_bit_width() {
32 => "__nac3_ndarray_calc_size",
64 => "__nac3_ndarray_calc_size64",
bw => unreachable!("Unsupported size type bit width: {}", bw),
};
let ndarray_calc_size_fn_t = llvm_usize.fn_type(
&[llvm_pusize.into(), llvm_usize.into(), llvm_usize.into(), llvm_usize.into()],
false,
);
let ndarray_calc_size_fn =
ctx.module.get_function(ndarray_calc_size_fn_name).unwrap_or_else(|| {
ctx.module.add_function(ndarray_calc_size_fn_name, ndarray_calc_size_fn_t, None)
});
// Get (or declare then get if does not exist) the corresponding function
ctx.module.get_function(&fn_name).unwrap_or_else(|| {
let fn_type = build_func_type();
ctx.module.add_function(&fn_name, fn_type, None)
})
}
fn get_ndarray_struct_ptr<'ctx>(ctx: &'ctx Context, size_type: IntType<'ctx>) -> PointerType<'ctx> {
let i8_type = ctx.i8_type();
let ndarray_ty = NpArrayType { size_type, elem_type: i8_type.as_basic_type_enum() };
let struct_ty = ndarray_ty.fields().whole_struct.as_struct_type(ctx);
struct_ty.ptr_type(AddressSpace::default())
}
pub fn call_nac3_ndarray_size<'ctx>(
ctx: &CodeGenContext<'ctx, '_>,
ndarray: NpArrayValue<'ctx>,
) -> IntValue<'ctx> {
let size_type = ndarray.ty.size_type;
let function = get_size_type_dependent_function(ctx, size_type, "__nac3_ndarray_size", || {
size_type.fn_type(&[get_ndarray_struct_ptr(ctx.ctx, size_type).into()], false)
});
let begin = begin.unwrap_or_else(|| llvm_usize.const_zero());
let end = end.unwrap_or_else(|| dims.size(ctx, generator));
ctx.builder
.build_call(
ndarray_calc_size_fn,
&[
dims.base_ptr(ctx, generator).into(),
dims.size(ctx, generator).into(),
begin.into(),
end.into(),
],
"",
)
.map(CallSiteValue::try_as_basic_value)
.map(|v| v.map_left(BasicValueEnum::into_int_value))
.map(Either::unwrap_left)
.build_call(function, &[ndarray.ptr.into()], "size")
.unwrap()
.try_as_basic_value()
.unwrap_left()
.into_int_value()
}
/// Generates a call to `__nac3_ndarray_calc_nd_indices`. Returns a [`TypeArrayLikeAdpater`]
/// containing `i32` indices of the flattened index.
///
/// * `index` - The index to compute the multidimensional index for.
/// * `ndarray` - LLVM pointer to the `NDArray`. This value must be the LLVM representation of an
/// `NDArray`.
pub fn call_ndarray_calc_nd_indices<'ctx, G: CodeGenerator + ?Sized>(
generator: &G,
ctx: &mut CodeGenContext<'ctx, '_>,
index: IntValue<'ctx>,
ndarray: NDArrayValue<'ctx>,
) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
let llvm_void = ctx.ctx.void_type();
let llvm_i32 = ctx.ctx.i32_type();
let llvm_usize = generator.get_size_type(ctx.ctx);
let llvm_pi32 = llvm_i32.ptr_type(AddressSpace::default());
let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
let ndarray_calc_nd_indices_fn_name = match llvm_usize.get_bit_width() {
32 => "__nac3_ndarray_calc_nd_indices",
64 => "__nac3_ndarray_calc_nd_indices64",
bw => unreachable!("Unsupported size type bit width: {}", bw),
};
let ndarray_calc_nd_indices_fn =
ctx.module.get_function(ndarray_calc_nd_indices_fn_name).unwrap_or_else(|| {
let fn_type = llvm_void.fn_type(
&[llvm_usize.into(), llvm_pusize.into(), llvm_usize.into(), llvm_pi32.into()],
false,
);
ctx.module.add_function(ndarray_calc_nd_indices_fn_name, fn_type, None)
});
let ndarray_num_dims = ndarray.load_ndims(ctx);
let ndarray_dims = ndarray.dim_sizes();
let indices = ctx.builder.build_array_alloca(llvm_i32, ndarray_num_dims, "").unwrap();
ctx.builder
.build_call(
ndarray_calc_nd_indices_fn,
&[
index.into(),
ndarray_dims.base_ptr(ctx, generator).into(),
ndarray_num_dims.into(),
indices.into(),
],
"",
)
.unwrap();
TypedArrayLikeAdapter::from(
ArraySliceValue::from_ptr_val(indices, ndarray_num_dims, None),
Box::new(|_, v| v.into_int_value()),
Box::new(|_, v| v.into()),
)
}
fn call_ndarray_flatten_index_impl<'ctx, G, Indices>(
generator: &G,
ctx: &CodeGenContext<'ctx, '_>,
ndarray: NDArrayValue<'ctx>,
indices: &Indices,
) -> IntValue<'ctx>
where
G: CodeGenerator + ?Sized,
Indices: ArrayLikeIndexer<'ctx>,
{
let llvm_i32 = ctx.ctx.i32_type();
let llvm_usize = generator.get_size_type(ctx.ctx);
let llvm_pi32 = llvm_i32.ptr_type(AddressSpace::default());
let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
debug_assert_eq!(
IntType::try_from(indices.element_type(ctx, generator))
.map(IntType::get_bit_width)
.unwrap_or_default(),
llvm_i32.get_bit_width(),
"Expected i32 value for argument `indices` to `call_ndarray_flatten_index_impl`"
);
debug_assert_eq!(
indices.size(ctx, generator).get_type().get_bit_width(),
llvm_usize.get_bit_width(),
"Expected usize integer value for argument `indices_size` to `call_ndarray_flatten_index_impl`"
);
let ndarray_flatten_index_fn_name = match llvm_usize.get_bit_width() {
32 => "__nac3_ndarray_flatten_index",
64 => "__nac3_ndarray_flatten_index64",
bw => unreachable!("Unsupported size type bit width: {}", bw),
};
let ndarray_flatten_index_fn =
ctx.module.get_function(ndarray_flatten_index_fn_name).unwrap_or_else(|| {
let fn_type = llvm_usize.fn_type(
&[llvm_pusize.into(), llvm_usize.into(), llvm_pi32.into(), llvm_usize.into()],
false,
);
ctx.module.add_function(ndarray_flatten_index_fn_name, fn_type, None)
});
let ndarray_num_dims = ndarray.load_ndims(ctx);
let ndarray_dims = ndarray.dim_sizes();
let index = ctx
.builder
.build_call(
ndarray_flatten_index_fn,
&[
ndarray_dims.base_ptr(ctx, generator).into(),
ndarray_num_dims.into(),
indices.base_ptr(ctx, generator).into(),
indices.size(ctx, generator).into(),
],
"",
)
.map(CallSiteValue::try_as_basic_value)
.map(|v| v.map_left(BasicValueEnum::into_int_value))
.map(Either::unwrap_left)
.unwrap();
index
}
/// Generates a call to `__nac3_ndarray_flatten_index`. Returns the flattened index for the
/// multidimensional index.
///
/// * `ndarray` - LLVM pointer to the `NDArray`. This value must be the LLVM representation of an
/// `NDArray`.
/// * `indices` - The multidimensional index to compute the flattened index for.
pub fn call_ndarray_flatten_index<'ctx, G, Index>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
ndarray: NDArrayValue<'ctx>,
indices: &Index,
) -> IntValue<'ctx>
where
G: CodeGenerator + ?Sized,
Index: ArrayLikeIndexer<'ctx>,
{
call_ndarray_flatten_index_impl(generator, ctx, ndarray, indices)
}
/// Generates a call to `__nac3_ndarray_calc_broadcast`. Returns a tuple containing the number of
/// dimension and size of each dimension of the resultant `ndarray`.
pub fn call_ndarray_calc_broadcast<'ctx, G: CodeGenerator + ?Sized>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
lhs: NDArrayValue<'ctx>,
rhs: NDArrayValue<'ctx>,
) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
let llvm_usize = generator.get_size_type(ctx.ctx);
let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
let ndarray_calc_broadcast_fn_name = match llvm_usize.get_bit_width() {
32 => "__nac3_ndarray_calc_broadcast",
64 => "__nac3_ndarray_calc_broadcast64",
bw => unreachable!("Unsupported size type bit width: {}", bw),
};
let ndarray_calc_broadcast_fn =
ctx.module.get_function(ndarray_calc_broadcast_fn_name).unwrap_or_else(|| {
let fn_type = llvm_usize.fn_type(
&[
llvm_pusize.into(),
llvm_usize.into(),
llvm_pusize.into(),
llvm_usize.into(),
llvm_pusize.into(),
],
false,
);
ctx.module.add_function(ndarray_calc_broadcast_fn_name, fn_type, None)
});
let lhs_ndims = lhs.load_ndims(ctx);
let rhs_ndims = rhs.load_ndims(ctx);
let min_ndims = llvm_intrinsics::call_int_umin(ctx, lhs_ndims, rhs_ndims, None);
gen_for_callback_incrementing(
generator,
ctx,
llvm_usize.const_zero(),
(min_ndims, false),
|generator, ctx, _, idx| {
let idx = ctx.builder.build_int_sub(min_ndims, idx, "").unwrap();
let (lhs_dim_sz, rhs_dim_sz) = unsafe {
(
lhs.dim_sizes().get_typed_unchecked(ctx, generator, &idx, None),
rhs.dim_sizes().get_typed_unchecked(ctx, generator, &idx, None),
)
};
let llvm_usize_const_one = llvm_usize.const_int(1, false);
let lhs_eqz = ctx
.builder
.build_int_compare(IntPredicate::EQ, lhs_dim_sz, llvm_usize_const_one, "")
.unwrap();
let rhs_eqz = ctx
.builder
.build_int_compare(IntPredicate::EQ, rhs_dim_sz, llvm_usize_const_one, "")
.unwrap();
let lhs_or_rhs_eqz = ctx.builder.build_or(lhs_eqz, rhs_eqz, "").unwrap();
let lhs_eq_rhs = ctx
.builder
.build_int_compare(IntPredicate::EQ, lhs_dim_sz, rhs_dim_sz, "")
.unwrap();
let is_compatible = ctx.builder.build_or(lhs_or_rhs_eqz, lhs_eq_rhs, "").unwrap();
ctx.make_assert(
generator,
is_compatible,
"0:ValueError",
"operands could not be broadcast together",
[None, None, None],
ctx.current_loc,
);
Ok(())
},
llvm_usize.const_int(1, false),
)
.unwrap();
let max_ndims = llvm_intrinsics::call_int_umax(ctx, lhs_ndims, rhs_ndims, None);
let lhs_dims = lhs.dim_sizes().base_ptr(ctx, generator);
let lhs_ndims = lhs.load_ndims(ctx);
let rhs_dims = rhs.dim_sizes().base_ptr(ctx, generator);
let rhs_ndims = rhs.load_ndims(ctx);
let out_dims = ctx.builder.build_array_alloca(llvm_usize, max_ndims, "").unwrap();
let out_dims = ArraySliceValue::from_ptr_val(out_dims, max_ndims, None);
ctx.builder
.build_call(
ndarray_calc_broadcast_fn,
&[
lhs_dims.into(),
lhs_ndims.into(),
rhs_dims.into(),
rhs_ndims.into(),
out_dims.base_ptr(ctx, generator).into(),
],
"",
)
.unwrap();
TypedArrayLikeAdapter::from(
out_dims,
Box::new(|_, v| v.into_int_value()),
Box::new(|_, v| v.into()),
)
}
/// Generates a call to `__nac3_ndarray_calc_broadcast_idx`. Returns an [`ArrayAllocaValue`]
/// containing the indices used for accessing `array` corresponding to the index of the broadcasted
/// array `broadcast_idx`.
pub fn call_ndarray_calc_broadcast_index<
'ctx,
G: CodeGenerator + ?Sized,
BroadcastIdx: UntypedArrayLikeAccessor<'ctx>,
>(
generator: &mut G,
ctx: &mut CodeGenContext<'ctx, '_>,
array: NDArrayValue<'ctx>,
broadcast_idx: &BroadcastIdx,
) -> TypedArrayLikeAdapter<'ctx, IntValue<'ctx>> {
let llvm_i32 = ctx.ctx.i32_type();
let llvm_usize = generator.get_size_type(ctx.ctx);
let llvm_pi32 = llvm_i32.ptr_type(AddressSpace::default());
let llvm_pusize = llvm_usize.ptr_type(AddressSpace::default());
let ndarray_calc_broadcast_fn_name = match llvm_usize.get_bit_width() {
32 => "__nac3_ndarray_calc_broadcast_idx",
64 => "__nac3_ndarray_calc_broadcast_idx64",
bw => unreachable!("Unsupported size type bit width: {}", bw),
};
let ndarray_calc_broadcast_fn =
ctx.module.get_function(ndarray_calc_broadcast_fn_name).unwrap_or_else(|| {
let fn_type = llvm_usize.fn_type(
&[llvm_pusize.into(), llvm_usize.into(), llvm_pi32.into(), llvm_pi32.into()],
false,
);
ctx.module.add_function(ndarray_calc_broadcast_fn_name, fn_type, None)
});
let broadcast_size = broadcast_idx.size(ctx, generator);
let out_idx = ctx.builder.build_array_alloca(llvm_i32, broadcast_size, "").unwrap();
let array_dims = array.dim_sizes().base_ptr(ctx, generator);
let array_ndims = array.load_ndims(ctx);
let broadcast_idx_ptr = unsafe {
broadcast_idx.ptr_offset_unchecked(ctx, generator, &llvm_usize.const_zero(), None)
};
ctx.builder
.build_call(
ndarray_calc_broadcast_fn,
&[array_dims.into(), array_ndims.into(), broadcast_idx_ptr.into(), out_idx.into()],
"",
)
.unwrap();
TypedArrayLikeAdapter::from(
ArraySliceValue::from_ptr_val(out_idx, broadcast_size, None),
Box::new(|_, v| v.into_int_value()),
Box::new(|_, v| v.into()),
)
}

641
nac3core/src/codegen/numpy.rs
View File

File diff suppressed because it is too large Load Diff

1
nac3core/src/lib.rs
View File

@ -23,3 +23,4 @@ pub mod codegen;
pub mod symbol_resolver;
pub mod toplevel;
pub mod typecheck;
pub mod util;

26
nac3core/src/toplevel/builtins.rs
View File

@ -1,5 +1,6 @@
use std::iter::once;
use crate::util::SizeVariant;
use helper::{debug_assert_prim_is_allowed, make_exception_fields, PrimDefDetails};
use indexmap::IndexMap;
use inkwell::{
@ -278,19 +279,10 @@ pub fn get_builtins(unifier: &mut Unifier, primitives: &PrimitiveStore) -> Built
.collect()
}
/// A helper enum used by [`BuiltinBuilder`]
#[derive(Clone, Copy)]
enum SizeVariant {
Bits32,
Bits64,
}
impl SizeVariant {
fn of_int(self, primitives: &PrimitiveStore) -> Type {
match self {
SizeVariant::Bits32 => primitives.int32,
SizeVariant::Bits64 => primitives.int64,
}
fn size_variant_to_int_type(variant: SizeVariant, primitives: &PrimitiveStore) -> Type {
match variant {
SizeVariant::Bits32 => primitives.int32,
SizeVariant::Bits64 => primitives.int64,
}
}
@ -1061,7 +1053,7 @@ impl<'a> BuiltinBuilder<'a> {
);
// The size variant of the function determines the size of the returned int.
let int_sized = size_variant.of_int(self.primitives);
let int_sized = size_variant_to_int_type(size_variant, self.primitives);
let ndarray_int_sized =
make_ndarray_ty(self.unifier, self.primitives, Some(int_sized), Some(common_ndim.ty));
@ -1086,7 +1078,7 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let ret_elem_ty = size_variant.of_int(&ctx.primitives);
let ret_elem_ty = size_variant_to_int_type(size_variant, &ctx.primitives);
Ok(Some(builtin_fns::call_round(generator, ctx, (arg_ty, arg), ret_elem_ty)?))
}),
)
@ -1127,7 +1119,7 @@ impl<'a> BuiltinBuilder<'a> {
make_ndarray_ty(self.unifier, self.primitives, Some(float), Some(common_ndim.ty));
// The size variant of the function determines the type of int returned
let int_sized = size_variant.of_int(self.primitives);
let int_sized = size_variant_to_int_type(size_variant, self.primitives);
let ndarray_int_sized =
make_ndarray_ty(self.unifier, self.primitives, Some(int_sized), Some(common_ndim.ty));
@ -1150,7 +1142,7 @@ impl<'a> BuiltinBuilder<'a> {
let arg_ty = fun.0.args[0].ty;
let arg = args[0].1.clone().to_basic_value_enum(ctx, generator, arg_ty)?;
let ret_elem_ty = size_variant.of_int(&ctx.primitives);
let ret_elem_ty = size_variant_to_int_type(size_variant, &ctx.primitives);
let func = match kind {
Kind::Ceil => builtin_fns::call_ceil,
Kind::Floor => builtin_fns::call_floor,

1
nac3core/src/toplevel/mod.rs
View File

@ -34,6 +34,7 @@ pub mod numpy;
pub mod type_annotation;
use composer::*;
use type_annotation::*;
#[cfg(test)]
mod test;

5
nac3core/src/util.rs Normal file
View File

@ -0,0 +1,5 @@
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum SizeVariant {
Bits32,
Bits64,
}