core: add irrt_test
This commit is contained in:
parent
87511ac749
commit
1e06a3d199
|
@ -13,6 +13,7 @@
|
|||
''
|
||||
mkdir -p $out/bin
|
||||
ln -s ${pkgs.llvmPackages_14.clang-unwrapped}/bin/clang $out/bin/clang-irrt
|
||||
ln -s ${pkgs.llvmPackages_14.clang}/bin/clang $out/bin/clang-irrt-test
|
||||
ln -s ${pkgs.llvmPackages_14.llvm.out}/bin/llvm-as $out/bin/llvm-as-irrt
|
||||
'';
|
||||
nac3artiq = pkgs.python3Packages.toPythonModule (
|
||||
|
@ -23,6 +24,7 @@
|
|||
cargoLock = {
|
||||
lockFile = ./Cargo.lock;
|
||||
};
|
||||
cargoTestFlags = [ "--features" "test" ];
|
||||
passthru.cargoLock = cargoLock;
|
||||
nativeBuildInputs = [ pkgs.python3 pkgs.llvmPackages_14.clang llvm-tools-irrt pkgs.llvmPackages_14.llvm.out llvm-nac3 ];
|
||||
buildInputs = [ pkgs.python3 llvm-nac3 ];
|
||||
|
|
|
@ -1,3 +1,6 @@
|
|||
[features]
|
||||
test = []
|
||||
|
||||
[package]
|
||||
name = "nac3core"
|
||||
version = "0.1.0"
|
||||
|
|
|
@ -7,8 +7,8 @@ use std::{
|
|||
process::{Command, Stdio},
|
||||
};
|
||||
|
||||
fn main() {
|
||||
const FILE: &str = "src/codegen/irrt/irrt.cpp";
|
||||
fn compile_irrt(irrt_dir: &Path, out_dir: &Path) {
|
||||
let irrt_cpp_path = irrt_dir.join("irrt.cpp");
|
||||
|
||||
/*
|
||||
* HACK: Sadly, clang doesn't let us emit generic LLVM bitcode.
|
||||
|
@ -16,7 +16,7 @@ fn main() {
|
|||
*/
|
||||
let flags: &[&str] = &[
|
||||
"--target=wasm32",
|
||||
FILE,
|
||||
irrt_cpp_path.to_str().unwrap(),
|
||||
"-x",
|
||||
"c++",
|
||||
"-fno-discard-value-names",
|
||||
|
@ -31,13 +31,13 @@ fn main() {
|
|||
"-S",
|
||||
"-Wall",
|
||||
"-Wextra",
|
||||
"-I",
|
||||
irrt_dir.to_str().unwrap(),
|
||||
"-o",
|
||||
"-",
|
||||
];
|
||||
|
||||
println!("cargo:rerun-if-changed={FILE}");
|
||||
let out_dir = env::var("OUT_DIR").unwrap();
|
||||
let out_path = Path::new(&out_dir);
|
||||
println!("cargo:rerun-if-changed={path}", path = irrt_cpp_path.to_str().unwrap());
|
||||
|
||||
let output = Command::new("clang-irrt")
|
||||
.args(flags)
|
||||
|
@ -65,18 +65,64 @@ fn main() {
|
|||
|
||||
println!("cargo:rerun-if-env-changed=DEBUG_DUMP_IRRT");
|
||||
if env::var("DEBUG_DUMP_IRRT").is_ok() {
|
||||
let mut file = File::create(out_path.join("irrt.ll")).unwrap();
|
||||
let mut file = File::create(out_dir.join("irrt.ll")).unwrap();
|
||||
file.write_all(output.as_bytes()).unwrap();
|
||||
let mut file = File::create(out_path.join("irrt-filtered.ll")).unwrap();
|
||||
let mut file = File::create(out_dir.join("irrt-filtered.ll")).unwrap();
|
||||
file.write_all(filtered_output.as_bytes()).unwrap();
|
||||
}
|
||||
|
||||
let mut llvm_as = Command::new("llvm-as-irrt")
|
||||
.stdin(Stdio::piped())
|
||||
.arg("-o")
|
||||
.arg(out_path.join("irrt.bc"))
|
||||
.arg(out_dir.join("irrt.bc"))
|
||||
.spawn()
|
||||
.unwrap();
|
||||
llvm_as.stdin.as_mut().unwrap().write_all(filtered_output.as_bytes()).unwrap();
|
||||
assert!(llvm_as.wait().unwrap().success());
|
||||
}
|
||||
|
||||
fn compile_irrt_test(irrt_dir: &Path, out_dir: &Path) {
|
||||
let irrt_test_cpp_path = irrt_dir.join("irrt_test.cpp");
|
||||
let exe_path = out_dir.join("irrt_test.out");
|
||||
|
||||
let flags: &[&str] = &[
|
||||
irrt_test_cpp_path.to_str().unwrap(),
|
||||
"-x",
|
||||
"c++",
|
||||
"-I",
|
||||
irrt_dir.to_str().unwrap(),
|
||||
"-g",
|
||||
"-fno-discard-value-names",
|
||||
"-O0",
|
||||
"-Wall",
|
||||
"-Wextra",
|
||||
"-lm", // for `tgamma()`, `lgamma()`
|
||||
"-o",
|
||||
exe_path.to_str().unwrap(),
|
||||
];
|
||||
|
||||
Command::new("clang-irrt-test")
|
||||
.args(flags)
|
||||
.output()
|
||||
.map(|o| {
|
||||
assert!(o.status.success(), "{}", std::str::from_utf8(&o.stderr).unwrap());
|
||||
o
|
||||
})
|
||||
.unwrap();
|
||||
println!("cargo:rerun-if-changed={path}", path = irrt_test_cpp_path.to_str().unwrap());
|
||||
}
|
||||
|
||||
fn main() {
|
||||
let out_dir = env::var("OUT_DIR").unwrap();
|
||||
let out_dir = Path::new(&out_dir);
|
||||
|
||||
let irrt_dir = Path::new("src/codegen/irrt");
|
||||
|
||||
compile_irrt(irrt_dir, out_dir);
|
||||
|
||||
// https://github.com/rust-lang/cargo/issues/2549
|
||||
// `cargo test -F test` to also build `irrt_test.cpp
|
||||
if cfg!(feature = "test") {
|
||||
compile_irrt_test(irrt_dir, out_dir);
|
||||
}
|
||||
}
|
||||
|
|
|
@ -1,428 +1,3 @@
|
|||
using int8_t = _BitInt(8);
|
||||
using uint8_t = unsigned _BitInt(8);
|
||||
using int32_t = _BitInt(32);
|
||||
using uint32_t = unsigned _BitInt(32);
|
||||
using int64_t = _BitInt(64);
|
||||
using uint64_t = unsigned _BitInt(64);
|
||||
#include "irrt.hpp"
|
||||
|
||||
// NDArray indices are always `uint32_t`.
|
||||
using NDIndex = uint32_t;
|
||||
// The type of an index or a value describing the length of a range/slice is always `int32_t`.
|
||||
using SliceIndex = int32_t;
|
||||
|
||||
namespace {
|
||||
template <typename T>
|
||||
const T& max(const T& a, const T& b) {
|
||||
return a > b ? a : b;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
const T& min(const T& a, const 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>
|
||||
T __nac3_int_exp_impl(T base, T exp) {
|
||||
T res = 1;
|
||||
/* repeated squaring method */
|
||||
do {
|
||||
if (exp & 1) {
|
||||
res *= base; /* for n odd */
|
||||
}
|
||||
exp >>= 1;
|
||||
base *= base;
|
||||
} while (exp);
|
||||
return res;
|
||||
}
|
||||
|
||||
template <typename SizeT>
|
||||
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>
|
||||
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>
|
||||
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>
|
||||
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>
|
||||
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];
|
||||
}
|
||||
}
|
||||
} // namespace
|
||||
|
||||
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) {\
|
||||
return __nac3_int_exp_impl(base, exp);\
|
||||
}
|
||||
|
||||
DEF_nac3_int_exp_(int32_t)
|
||||
DEF_nac3_int_exp_(int64_t)
|
||||
DEF_nac3_int_exp_(uint32_t)
|
||||
DEF_nac3_int_exp_(uint64_t)
|
||||
|
||||
SliceIndex __nac3_slice_index_bound(SliceIndex i, const SliceIndex len) {
|
||||
if (i < 0) {
|
||||
i = len + i;
|
||||
}
|
||||
if (i < 0) {
|
||||
return 0;
|
||||
} else if (i > len) {
|
||||
return len;
|
||||
}
|
||||
return i;
|
||||
}
|
||||
|
||||
SliceIndex __nac3_range_slice_len(
|
||||
const SliceIndex start,
|
||||
const SliceIndex end,
|
||||
const SliceIndex step
|
||||
) {
|
||||
SliceIndex diff = end - start;
|
||||
if (diff > 0 && step > 0) {
|
||||
return ((diff - 1) / step) + 1;
|
||||
} else if (diff < 0 && step < 0) {
|
||||
return ((diff + 1) / step) + 1;
|
||||
} else {
|
||||
return 0;
|
||||
}
|
||||
}
|
||||
|
||||
// Handle list assignment and dropping part of the list when
|
||||
// both dest_step and src_step are +1.
|
||||
// - All the index must *not* be out-of-bound or negative,
|
||||
// - The end index is *inclusive*,
|
||||
// - The length of src and dest slice size should already
|
||||
// be checked: if dest.step == 1 then len(src) <= len(dest) else len(src) == len(dest)
|
||||
SliceIndex __nac3_list_slice_assign_var_size(
|
||||
SliceIndex dest_start,
|
||||
SliceIndex dest_end,
|
||||
SliceIndex dest_step,
|
||||
uint8_t* dest_arr,
|
||||
SliceIndex dest_arr_len,
|
||||
SliceIndex src_start,
|
||||
SliceIndex src_end,
|
||||
SliceIndex src_step,
|
||||
uint8_t* src_arr,
|
||||
SliceIndex src_arr_len,
|
||||
const SliceIndex size
|
||||
) {
|
||||
/* if dest_arr_len == 0, do nothing since we do not support extending list */
|
||||
if (dest_arr_len == 0) return dest_arr_len;
|
||||
/* if both step is 1, memmove directly, handle the dropping of the list, and shrink size */
|
||||
if (src_step == dest_step && dest_step == 1) {
|
||||
const SliceIndex src_len = (src_end >= src_start) ? (src_end - src_start + 1) : 0;
|
||||
const SliceIndex dest_len = (dest_end >= dest_start) ? (dest_end - dest_start + 1) : 0;
|
||||
if (src_len > 0) {
|
||||
__builtin_memmove(
|
||||
dest_arr + dest_start * size,
|
||||
src_arr + src_start * size,
|
||||
src_len * size
|
||||
);
|
||||
}
|
||||
if (dest_len > 0) {
|
||||
/* dropping */
|
||||
__builtin_memmove(
|
||||
dest_arr + (dest_start + src_len) * size,
|
||||
dest_arr + (dest_end + 1) * size,
|
||||
(dest_arr_len - dest_end - 1) * size
|
||||
);
|
||||
}
|
||||
/* shrink size */
|
||||
return dest_arr_len - (dest_len - src_len);
|
||||
}
|
||||
/* if two range overlaps, need alloca */
|
||||
uint8_t need_alloca =
|
||||
(dest_arr == src_arr)
|
||||
&& !(
|
||||
max(dest_start, dest_end) < min(src_start, src_end)
|
||||
|| max(src_start, src_end) < min(dest_start, dest_end)
|
||||
);
|
||||
if (need_alloca) {
|
||||
uint8_t* tmp = reinterpret_cast<uint8_t *>(__builtin_alloca(src_arr_len * size));
|
||||
__builtin_memcpy(tmp, src_arr, src_arr_len * size);
|
||||
src_arr = tmp;
|
||||
}
|
||||
SliceIndex src_ind = src_start;
|
||||
SliceIndex dest_ind = dest_start;
|
||||
for (;
|
||||
(src_step > 0) ? (src_ind <= src_end) : (src_ind >= src_end);
|
||||
src_ind += src_step, dest_ind += dest_step
|
||||
) {
|
||||
/* for constant optimization */
|
||||
if (size == 1) {
|
||||
__builtin_memcpy(dest_arr + dest_ind, src_arr + src_ind, 1);
|
||||
} else if (size == 4) {
|
||||
__builtin_memcpy(dest_arr + dest_ind * 4, src_arr + src_ind * 4, 4);
|
||||
} else if (size == 8) {
|
||||
__builtin_memcpy(dest_arr + dest_ind * 8, src_arr + src_ind * 8, 8);
|
||||
} else {
|
||||
/* memcpy for var size, cannot overlap after previous alloca */
|
||||
__builtin_memcpy(dest_arr + dest_ind * size, src_arr + src_ind * size, size);
|
||||
}
|
||||
}
|
||||
/* only dest_step == 1 can we shrink the dest list. */
|
||||
/* size should be ensured prior to calling this function */
|
||||
if (dest_step == 1 && dest_end >= dest_start) {
|
||||
__builtin_memmove(
|
||||
dest_arr + dest_ind * size,
|
||||
dest_arr + (dest_end + 1) * size,
|
||||
(dest_arr_len - dest_end - 1) * size
|
||||
);
|
||||
return dest_arr_len - (dest_end - dest_ind) - 1;
|
||||
}
|
||||
return dest_arr_len;
|
||||
}
|
||||
|
||||
int32_t __nac3_isinf(double x) {
|
||||
return __builtin_isinf(x);
|
||||
}
|
||||
|
||||
int32_t __nac3_isnan(double x) {
|
||||
return __builtin_isnan(x);
|
||||
}
|
||||
|
||||
double tgamma(double arg);
|
||||
|
||||
double __nac3_gamma(double z) {
|
||||
// Handling for denormals
|
||||
// | x | Python gamma(x) | C tgamma(x) |
|
||||
// --- | ----------------- | --------------- | ----------- |
|
||||
// (1) | nan | nan | nan |
|
||||
// (2) | -inf | -inf | inf |
|
||||
// (3) | inf | inf | inf |
|
||||
// (4) | 0.0 | inf | inf |
|
||||
// (5) | {-1.0, -2.0, ...} | inf | nan |
|
||||
|
||||
// (1)-(3)
|
||||
if (__builtin_isinf(z) || __builtin_isnan(z)) {
|
||||
return z;
|
||||
}
|
||||
|
||||
double v = tgamma(z);
|
||||
|
||||
// (4)-(5)
|
||||
return __builtin_isinf(v) || __builtin_isnan(v) ? __builtin_inf() : v;
|
||||
}
|
||||
|
||||
double lgamma(double arg);
|
||||
|
||||
double __nac3_gammaln(double x) {
|
||||
// libm's handling of value overflows differs from scipy:
|
||||
// - scipy: gammaln(-inf) -> -inf
|
||||
// - libm : lgamma(-inf) -> inf
|
||||
|
||||
if (__builtin_isinf(x)) {
|
||||
return x;
|
||||
}
|
||||
|
||||
return lgamma(x);
|
||||
}
|
||||
|
||||
double j0(double x);
|
||||
|
||||
double __nac3_j0(double x) {
|
||||
// libm's handling of value overflows differs from scipy:
|
||||
// - scipy: j0(inf) -> nan
|
||||
// - libm : j0(inf) -> 0.0
|
||||
|
||||
if (__builtin_isinf(x)) {
|
||||
return __builtin_nan("");
|
||||
}
|
||||
|
||||
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);
|
||||
}
|
||||
} // extern "C"
|
||||
// All the implementations are from `irrt.hpp`
|
||||
|
|
|
@ -0,0 +1,437 @@
|
|||
#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
|
||||
|
||||
// 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>
|
||||
static T __nac3_int_exp_impl(T base, T exp) {
|
||||
T res = 1;
|
||||
/* repeated squaring method */
|
||||
do {
|
||||
if (exp & 1) {
|
||||
res *= base; /* for n odd */
|
||||
}
|
||||
exp >>= 1;
|
||||
base *= base;
|
||||
} while (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) {\
|
||||
return __nac3_int_exp_impl(base, exp);\
|
||||
}
|
||||
|
||||
DEF_nac3_int_exp_(int32_t)
|
||||
DEF_nac3_int_exp_(int64_t)
|
||||
DEF_nac3_int_exp_(uint32_t)
|
||||
DEF_nac3_int_exp_(uint64_t)
|
||||
|
||||
SliceIndex __nac3_slice_index_bound(SliceIndex i, const SliceIndex len) {
|
||||
if (i < 0) {
|
||||
i = len + i;
|
||||
}
|
||||
if (i < 0) {
|
||||
return 0;
|
||||
} else if (i > len) {
|
||||
return len;
|
||||
}
|
||||
return i;
|
||||
}
|
||||
|
||||
SliceIndex __nac3_range_slice_len(
|
||||
const SliceIndex start,
|
||||
const SliceIndex end,
|
||||
const SliceIndex step
|
||||
) {
|
||||
SliceIndex diff = end - start;
|
||||
if (diff > 0 && step > 0) {
|
||||
return ((diff - 1) / step) + 1;
|
||||
} else if (diff < 0 && step < 0) {
|
||||
return ((diff + 1) / step) + 1;
|
||||
} else {
|
||||
return 0;
|
||||
}
|
||||
}
|
||||
|
||||
// Handle list assignment and dropping part of the list when
|
||||
// both dest_step and src_step are +1.
|
||||
// - All the index must *not* be out-of-bound or negative,
|
||||
// - The end index is *inclusive*,
|
||||
// - The length of src and dest slice size should already
|
||||
// be checked: if dest.step == 1 then len(src) <= len(dest) else len(src) == len(dest)
|
||||
SliceIndex __nac3_list_slice_assign_var_size(
|
||||
SliceIndex dest_start,
|
||||
SliceIndex dest_end,
|
||||
SliceIndex dest_step,
|
||||
uint8_t *dest_arr,
|
||||
SliceIndex dest_arr_len,
|
||||
SliceIndex src_start,
|
||||
SliceIndex src_end,
|
||||
SliceIndex src_step,
|
||||
uint8_t *src_arr,
|
||||
SliceIndex src_arr_len,
|
||||
const SliceIndex size
|
||||
) {
|
||||
/* if dest_arr_len == 0, do nothing since we do not support extending list */
|
||||
if (dest_arr_len == 0) return dest_arr_len;
|
||||
/* if both step is 1, memmove directly, handle the dropping of the list, and shrink size */
|
||||
if (src_step == dest_step && dest_step == 1) {
|
||||
const SliceIndex src_len = (src_end >= src_start) ? (src_end - src_start + 1) : 0;
|
||||
const SliceIndex dest_len = (dest_end >= dest_start) ? (dest_end - dest_start + 1) : 0;
|
||||
if (src_len > 0) {
|
||||
__builtin_memmove(
|
||||
dest_arr + dest_start * size,
|
||||
src_arr + src_start * size,
|
||||
src_len * size
|
||||
);
|
||||
}
|
||||
if (dest_len > 0) {
|
||||
/* dropping */
|
||||
__builtin_memmove(
|
||||
dest_arr + (dest_start + src_len) * size,
|
||||
dest_arr + (dest_end + 1) * size,
|
||||
(dest_arr_len - dest_end - 1) * size
|
||||
);
|
||||
}
|
||||
/* shrink size */
|
||||
return dest_arr_len - (dest_len - src_len);
|
||||
}
|
||||
/* if two range overlaps, need alloca */
|
||||
uint8_t need_alloca =
|
||||
(dest_arr == src_arr)
|
||||
&& !(
|
||||
max(dest_start, dest_end) < min(src_start, src_end)
|
||||
|| max(src_start, src_end) < min(dest_start, dest_end)
|
||||
);
|
||||
if (need_alloca) {
|
||||
uint8_t *tmp = reinterpret_cast<uint8_t *>(__builtin_alloca(src_arr_len * size));
|
||||
__builtin_memcpy(tmp, src_arr, src_arr_len * size);
|
||||
src_arr = tmp;
|
||||
}
|
||||
SliceIndex src_ind = src_start;
|
||||
SliceIndex dest_ind = dest_start;
|
||||
for (;
|
||||
(src_step > 0) ? (src_ind <= src_end) : (src_ind >= src_end);
|
||||
src_ind += src_step, dest_ind += dest_step
|
||||
) {
|
||||
/* for constant optimization */
|
||||
if (size == 1) {
|
||||
__builtin_memcpy(dest_arr + dest_ind, src_arr + src_ind, 1);
|
||||
} else if (size == 4) {
|
||||
__builtin_memcpy(dest_arr + dest_ind * 4, src_arr + src_ind * 4, 4);
|
||||
} else if (size == 8) {
|
||||
__builtin_memcpy(dest_arr + dest_ind * 8, src_arr + src_ind * 8, 8);
|
||||
} else {
|
||||
/* memcpy for var size, cannot overlap after previous alloca */
|
||||
__builtin_memcpy(dest_arr + dest_ind * size, src_arr + src_ind * size, size);
|
||||
}
|
||||
}
|
||||
/* only dest_step == 1 can we shrink the dest list. */
|
||||
/* size should be ensured prior to calling this function */
|
||||
if (dest_step == 1 && dest_end >= dest_start) {
|
||||
__builtin_memmove(
|
||||
dest_arr + dest_ind * size,
|
||||
dest_arr + (dest_end + 1) * size,
|
||||
(dest_arr_len - dest_end - 1) * size
|
||||
);
|
||||
return dest_arr_len - (dest_end - dest_ind) - 1;
|
||||
}
|
||||
return dest_arr_len;
|
||||
}
|
||||
|
||||
int32_t __nac3_isinf(double x) {
|
||||
return __builtin_isinf(x);
|
||||
}
|
||||
|
||||
int32_t __nac3_isnan(double x) {
|
||||
return __builtin_isnan(x);
|
||||
}
|
||||
|
||||
double tgamma(double arg);
|
||||
|
||||
double __nac3_gamma(double z) {
|
||||
// Handling for denormals
|
||||
// | x | Python gamma(x) | C tgamma(x) |
|
||||
// --- | ----------------- | --------------- | ----------- |
|
||||
// (1) | nan | nan | nan |
|
||||
// (2) | -inf | -inf | inf |
|
||||
// (3) | inf | inf | inf |
|
||||
// (4) | 0.0 | inf | inf |
|
||||
// (5) | {-1.0, -2.0, ...} | inf | nan |
|
||||
|
||||
// (1)-(3)
|
||||
if (__builtin_isinf(z) || __builtin_isnan(z)) {
|
||||
return z;
|
||||
}
|
||||
|
||||
double v = tgamma(z);
|
||||
|
||||
// (4)-(5)
|
||||
return __builtin_isinf(v) || __builtin_isnan(v) ? __builtin_inf() : v;
|
||||
}
|
||||
|
||||
double lgamma(double arg);
|
||||
|
||||
double __nac3_gammaln(double x) {
|
||||
// libm's handling of value overflows differs from scipy:
|
||||
// - scipy: gammaln(-inf) -> -inf
|
||||
// - libm : lgamma(-inf) -> inf
|
||||
|
||||
if (__builtin_isinf(x)) {
|
||||
return x;
|
||||
}
|
||||
|
||||
return lgamma(x);
|
||||
}
|
||||
|
||||
double j0(double x);
|
||||
|
||||
double __nac3_j0(double x) {
|
||||
// libm's handling of value overflows differs from scipy:
|
||||
// - scipy: j0(inf) -> nan
|
||||
// - libm : j0(inf) -> 0.0
|
||||
|
||||
if (__builtin_isinf(x)) {
|
||||
return __builtin_nan("");
|
||||
}
|
||||
|
||||
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);
|
||||
}
|
||||
}
|
|
@ -0,0 +1,62 @@
|
|||
#include <cstdint>
|
||||
#include <cstdio>
|
||||
#include <cstdlib>
|
||||
|
||||
#define IRRT_DONT_TYPEDEF_INTS
|
||||
#include "irrt.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);
|
||||
exit(1);
|
||||
}
|
||||
|
||||
#define test_fail() __test_fail(__FILE__, __LINE__);
|
||||
|
||||
template <typename T>
|
||||
bool arrays_match(int len, T *as, T *bs) {
|
||||
for (int i = 0; i < len; i++) {
|
||||
if (as[i] != bs[i]) return false;
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
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");
|
||||
}
|
||||
|
||||
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) {
|
||||
printf("expected %s: ", label);
|
||||
debug_print_array(format, len, expected);
|
||||
printf("got %s: ", label);
|
||||
debug_print_array(format, len, got);
|
||||
}
|
||||
|
||||
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);
|
||||
|
||||
uint64_t expected_strides[ndims] = { 3*5*7, 5*7, 7, 1 };
|
||||
if (!assert_arrays_match("strides", "%u", ndims, expected_strides, strides)) test_fail();
|
||||
}
|
||||
|
||||
int main() {
|
||||
test_strides_from_shape();
|
||||
return 0;
|
||||
}
|
|
@ -1,5 +1,7 @@
|
|||
use crate::typecheck::typedef::Type;
|
||||
|
||||
mod test;
|
||||
|
||||
use super::{
|
||||
classes::{
|
||||
ArrayLikeIndexer, ArrayLikeValue, ArraySliceValue, ListValue, NDArrayValue,
|
||||
|
|
|
@ -0,0 +1,26 @@
|
|||
#[cfg(test)]
|
||||
mod tests {
|
||||
use std::{path::Path, process::Command};
|
||||
|
||||
#[test]
|
||||
fn run_irrt_test() {
|
||||
assert!(
|
||||
cfg!(feature = "test"),
|
||||
"Please do `cargo test -F test` to compile `irrt_test.out` and run test"
|
||||
);
|
||||
|
||||
let irrt_test_out_path = Path::new(concat!(env!("OUT_DIR"), "/irrt_test.out"));
|
||||
let output = Command::new(irrt_test_out_path.to_str().unwrap()).output().unwrap();
|
||||
|
||||
if !output.status.success() {
|
||||
eprintln!("irrt_test failed with status {}:", output.status);
|
||||
eprintln!("====== stdout ======");
|
||||
eprintln!("{}", String::from_utf8(output.stdout).unwrap());
|
||||
eprintln!("====== stderr ======");
|
||||
eprintln!("{}", String::from_utf8(output.stderr).unwrap());
|
||||
eprintln!("====================");
|
||||
|
||||
panic!("irrt_test failed");
|
||||
}
|
||||
}
|
||||
}
|
|
@ -81,6 +81,7 @@ in rec {
|
|||
''
|
||||
mkdir -p $out/bin
|
||||
ln -s ${llvm-nac3}/bin/clang.exe $out/bin/clang-irrt.exe
|
||||
ln -s ${llvm-nac3}/bin/clang.exe $out/bin/clang-irrt-test.exe
|
||||
ln -s ${llvm-nac3}/bin/llvm-as.exe $out/bin/llvm-as-irrt.exe
|
||||
'';
|
||||
nac3artiq = pkgs.rustPlatform.buildRustPackage {
|
||||
|
|
Loading…
Reference in New Issue