Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

enh: remove simd_nightly feature flag for aarch64 #63

Merged
merged 18 commits into from
Mar 3, 2024
Merged

enh: remove simd_nightly feature flag for aarch64 #63

Show file tree
Hide file tree
Changes from 15 commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
51d821d
remove simd_nightly feature flag for aarch64 for i16
jvdd Feb 21, 2024
468c83e
add aarch64 target_arch to imports
jvdd Feb 21, 2024
03c0863
add aarch64 target_arch to imports in generic.rs
jvdd Feb 21, 2024
03b03e1
remove nightly_simd flag for impl_SIMDArgMinMax macro
jvdd Feb 21, 2024
6f48cc7
remove nightly_simd feature flag for aarch64 for i8 and i32
jvdd Feb 21, 2024
098661e
wrap up ints and uints
jvdd Feb 22, 2024
d3e19fd
add f32 support
jvdd Feb 23, 2024
641b1b0
fix arm
jvdd Feb 23, 2024
1e12aa6
:see_no_evil:
jvdd Feb 23, 2024
81842be
update other float types
jvdd Feb 24, 2024
75d619d
restrict cfg to arm and nightly_simd for floats
jvdd Feb 24, 2024
d575a56
restrict imports to arm & nightly_simd
jvdd Feb 24, 2024
773f08d
:recycle:
jvdd Feb 24, 2024
15ac377
build aarch64 as stable
jvdd Feb 24, 2024
ecc035f
use scalar implementation for 64 bit integers on aarch64
jvdd Feb 28, 2024
aa4220b
:broom: review code
jvdd Feb 29, 2024
cacfb3c
:broom: remove fixed TODO
jvdd Mar 1, 2024
4967e1b
:lotus_position: review code
jvdd Mar 1, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions Cargo.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,11 +14,11 @@ exclude = [".git*", "dev_utils/**/*", "tests/**/*"]


[dependencies]
num-traits = { version = "0.2.17", default-features = false }
num-traits = { version = "0.2", default-features = false }
half = { version = "2.3.1", default-features = false, features=["num-traits"], optional = true }
ndarray = { version = "0.15.6", default-features = false, optional = true}
arrow = { version = ">0", default-features = false, optional = true}
arrow2 = { version = ">0.0", default-features = false, optional = true}
ndarray = { version = "0.15.6", default-features = false, optional = true }
arrow = { version = ">0", default-features = false, optional = true }
arrow2 = { version = ">0.0", default-features = false, optional = true }
# once_cell = "1.16.0"

[features]
Expand Down
101 changes: 19 additions & 82 deletions src/lib.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -118,8 +118,8 @@ pub(crate) use simd::AVX512;
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
pub(crate) use simd::{SIMDArgMinMax, AVX2, SSE};
#[cfg(any(
all(target_arch = "aarch64", feature = "float"), // is stable for f64
all(any(target_arch = "arm", target_arch = "aarch64"), feature = "nightly_simd"),
all(target_arch = "arm", feature = "nightly_simd"),
target_arch = "aarch64"
))]
pub(crate) use simd::{SIMDArgMinMax, NEON};

Expand Down Expand Up @@ -304,9 +304,9 @@ macro_rules! impl_argminmax_int {
return unsafe { SSE::<Int>::argminmax(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$int_type>::NB_BITS < 64) {
if std::arch::is_aarch64_feature_detected!("neon") & (<$int_type>::NB_BITS < 64) {
jvdd marked this conversation as resolved.
Show resolved Hide resolved
// Scalar is faster for 64-bit numbers
return unsafe { NEON::<Int>::argminmax(self) }
}
Expand Down Expand Up @@ -349,7 +349,7 @@ macro_rules! impl_argminmax_int {
return unsafe { SSE::<Int>::argmin(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<Int>::argmin(self) }
Expand Down Expand Up @@ -392,7 +392,7 @@ macro_rules! impl_argminmax_int {
return unsafe { SSE::<Int>::argmax(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<Int>::argmax(self) }
Expand Down Expand Up @@ -443,10 +443,9 @@ macro_rules! impl_argminmax_float {
return unsafe { SSE::<FloatIgnoreNaN>::argminmax(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$float_type>::NB_BITS < 64) {
// NEON f64 is part of stable Rust (see code below this macro)
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatIgnoreNaN>::argminmax(self) }
}
}
Expand Down Expand Up @@ -483,10 +482,9 @@ macro_rules! impl_argminmax_float {
return unsafe { SSE::<FloatIgnoreNaN>::argmin(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$float_type>::NB_BITS < 64) {
// NEON f64 is part of stable Rust (see code below this macro)
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatIgnoreNaN>::argmin(self) }
}
}
Expand Down Expand Up @@ -523,10 +521,9 @@ macro_rules! impl_argminmax_float {
return unsafe { SSE::<FloatIgnoreNaN>::argmax(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$float_type>::NB_BITS < 64) {
// NEON f64 is part of stable Rust (see code below this macro)
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatIgnoreNaN>::argmax(self) }
}
}
Expand Down Expand Up @@ -563,10 +560,9 @@ macro_rules! impl_argminmax_float {
return unsafe { SSE::<FloatReturnNaN>::argminmax(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$float_type>::NB_BITS < 64) {
// We miss some NEON instructions for 64-bit numbers
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatReturnNaN>::argminmax(self) }
}
}
Expand Down Expand Up @@ -601,10 +597,9 @@ macro_rules! impl_argminmax_float {
return unsafe { SSE::<FloatReturnNaN>::argmin(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$float_type>::NB_BITS < 64) {
// We miss some NEON instructions for 64-bit numbers
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatReturnNaN>::argmin(self) }
}
}
Expand Down Expand Up @@ -639,10 +634,9 @@ macro_rules! impl_argminmax_float {
return unsafe { SSE::<FloatReturnNaN>::argmax(self) }
}
}
#[cfg(all(target_arch = "aarch64", feature = "nightly_simd"))]
#[cfg(target_arch = "aarch64")]
{
if std::arch::is_aarch64_feature_detected!("neon") & (<$float_type>::NB_BITS < 64) {
// We miss some NEON instructions for 64-bit numbers
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatReturnNaN>::argmax(self) }
}
}
Expand All @@ -660,68 +654,11 @@ macro_rules! impl_argminmax_float {
};
}

/// Implement ArgMinMax for &[f64] on aarch64 as NEON intrinsics for f64
/// are part of stable Rust on aarch64.
// Note: implementing this in a distinct impl block seemed more clean than
// hacking with unimpl_ macros in the simd_.rs files to avoid complaints
// from the compiler..
#[cfg(all(feature = "float", target_arch = "aarch64"))]
impl ArgMinMax for &[f64] {
fn argminmax(&self) -> (usize, usize) {
unsafe { NEON::<FloatIgnoreNaN>::argminmax(self) }
}
fn argmin(&self) -> usize {
unsafe { NEON::<FloatIgnoreNaN>::argmin(self) }
}
fn argmax(&self) -> usize {
unsafe { NEON::<FloatIgnoreNaN>::argmax(self) }
}
}

/// Implement NaNArgMinMax for &[f64] on aarch64 - the required intrinsics
/// for return nan are not part of stable Rust.
// Note: implementing this in a distinct impl block seemed more clean than
// hacking with unimpl_ macros in the simd_.rs files to avoid complaints
// from the compiler..
#[cfg(all(feature = "float", target_arch = "aarch64"))]
impl NaNArgMinMax for &[f64] {
fn nanargminmax(&self) -> (usize, usize) {
#[cfg(feature = "nightly_simd")]
{
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatReturnNaN>::argminmax(self) };
}
}
SCALAR::<FloatReturnNaN>::argminmax(self)
}
fn nanargmin(&self) -> usize {
#[cfg(feature = "nightly_simd")]
{
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatReturnNaN>::argmin(self) };
}
}
SCALAR::<FloatReturnNaN>::argmin(self)
}
fn nanargmax(&self) -> usize {
#[cfg(feature = "nightly_simd")]
{
if std::arch::is_aarch64_feature_detected!("neon") {
return unsafe { NEON::<FloatReturnNaN>::argmax(self) };
}
}
SCALAR::<FloatReturnNaN>::argmax(self)
}
}

// Implement ArgMinMax for (non-optional) integer rust primitive types
impl_argminmax_int!(i8, i16, i32, i64, u8, u16, u32, u64);
// Implement for (optional) float rust primitive types
#[cfg(all(feature = "float", not(target_arch = "aarch64")))]
#[cfg(feature = "float")]
impl_argminmax_float!(f32, f64);
// For aarch64 f64 is implemented in the two impl blocks above
#[cfg(all(feature = "float", target_arch = "aarch64"))]
impl_argminmax_float!(f32);

// Implement ArgMinMax for other data types
#[cfg(feature = "half")]
Expand Down
40 changes: 30 additions & 10 deletions src/simd/generic.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,9 +111,9 @@ where
/// - `impl_SIMDInit_Int!`
/// - called in the `simd_i*.rs` files
/// - called in the `simd_u*.rs` files
/// - `impl_SIMDInit_FloatIgnoreNaN!`
/// - see the `simd_f*_return_nan.rs` files
/// - `impl_SIMDInit_FloatReturnNaN!`
/// - see the `simd_f*_return_nan.rs` files
/// - `impl_SIMDInit_FloatIgnoreNaN!`
/// - see the `simd_f*_ignore_nan.rs` files
///
/// The current (default) implementation is for the Int case - see `impl_SIMDInit_Int!`
Expand Down Expand Up @@ -181,7 +181,12 @@ where

// --------------- Int (signed and unsigned)

#[cfg(any(target_arch = "x86", target_arch = "x86_64", feature = "nightly_simd"))]
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
target_arch = "aarch64",
feature = "nightly_simd"
))]
macro_rules! impl_SIMDInit_Int {
($scalar_dtype:ty, $simd_vec_dtype:ty, $simd_mask_dtype:ty, $lane_size:expr, $simd_struct:ty) => {
impl SIMDInit<$scalar_dtype, $simd_vec_dtype, $simd_mask_dtype, $lane_size>
Expand All @@ -192,13 +197,23 @@ macro_rules! impl_SIMDInit_Int {
};
}

#[cfg(any(target_arch = "x86", target_arch = "x86_64", feature = "nightly_simd"))]
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
target_arch = "aarch64",
feature = "nightly_simd"
))]
pub(crate) use impl_SIMDInit_Int; // Now classic paths Just Work™

// --------------- Float Return NaNs

#[cfg(any(feature = "float", feature = "half"))]
#[cfg(any(target_arch = "x86", target_arch = "x86_64", feature = "nightly_simd"))]
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
target_arch = "aarch64",
feature = "nightly_simd"
))]
macro_rules! impl_SIMDInit_FloatReturnNaN {
($scalar_dtype:ty, $simd_vec_dtype:ty, $simd_mask_dtype:ty, $lane_size:expr, $simd_struct:ty) => {
impl SIMDInit<$scalar_dtype, $simd_vec_dtype, $simd_mask_dtype, $lane_size>
Expand All @@ -221,7 +236,12 @@ macro_rules! impl_SIMDInit_FloatReturnNaN {
}

#[cfg(any(feature = "float", feature = "half"))]
#[cfg(any(target_arch = "x86", target_arch = "x86_64", feature = "nightly_simd"))]
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
target_arch = "aarch64",
feature = "nightly_simd"
))]
pub(crate) use impl_SIMDInit_FloatReturnNaN; // Now classic paths Just Work™

// --------------- Float Ignore NaNs
Expand All @@ -230,7 +250,7 @@ pub(crate) use impl_SIMDInit_FloatReturnNaN; // Now classic paths Just Work™
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
all(target_arch = "aarch64", feature = "float"), // is stable for f64
target_arch = "aarch64", // is stable for f64
feature = "nightly_simd"
))]
macro_rules! impl_SIMDInit_FloatIgnoreNaN {
Expand Down Expand Up @@ -306,7 +326,7 @@ macro_rules! impl_SIMDInit_FloatIgnoreNaN {
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
all(target_arch = "aarch64", feature = "float"), // is stable for f64
target_arch = "aarch64", // is stable for f64
feature = "nightly_simd"
))]
pub(crate) use impl_SIMDInit_FloatIgnoreNaN; // Now classic paths Just Work™
Expand Down Expand Up @@ -768,7 +788,7 @@ where
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
all(target_arch = "aarch64", feature = "float"), // is stable for f64
target_arch = "aarch64",
feature = "nightly_simd"
))]
macro_rules! impl_SIMDArgMinMax {
Expand Down Expand Up @@ -806,7 +826,7 @@ macro_rules! impl_SIMDArgMinMax {
#[cfg(any(
target_arch = "x86",
target_arch = "x86_64",
all(target_arch = "aarch64", feature = "float"), // is stable for f64
target_arch = "aarch64",
feature = "nightly_simd"
))]
pub(crate) use impl_SIMDArgMinMax; // Now classic paths Just Work™
Expand Down
Loading
Loading