Skip to content

NVPTX: Add f16 SIMD intrinsics #1626

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.

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

Merged
merged 1 commit into from
Aug 19, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
5 changes: 5 additions & 0 deletions crates/core_arch/src/nvptx/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,11 @@

use crate::ffi::c_void;

mod packed;

#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub use packed::*;

#[allow(improper_ctypes)]
extern "C" {
#[link_name = "llvm.nvvm.barrier0"]
Expand Down
143 changes: 143 additions & 0 deletions crates/core_arch/src/nvptx/packed.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,143 @@
//! NVPTX Packed data types (SIMD)
//!
//! Packed Data Types is what PTX calls SIMD types. See [PTX ISA (Packed Data Types)](https://docs.nvidia.com/cuda/parallel-thread-execution/#packed-data-types) for a full reference.

// Note: #[assert_instr] tests are not actually being run on nvptx due to being a `no_std` target incapable of running tests. Something like FileCheck would be appropriate for verifying the correct instruction is used.

use crate::intrinsics::simd::*;

#[allow(improper_ctypes)]
extern "C" {
#[link_name = "llvm.minnum.v2f16"]
fn llvm_f16x2_minnum(a: f16x2, b: f16x2) -> f16x2;
#[link_name = "llvm.minimum.v2f16"]
fn llvm_f16x2_minimum(a: f16x2, b: f16x2) -> f16x2;
#[link_name = "llvm.maxnum.v2f16"]
fn llvm_f16x2_maxnum(a: f16x2, b: f16x2) -> f16x2;
#[link_name = "llvm.maximum.v2f16"]
fn llvm_f16x2_maximum(a: f16x2, b: f16x2) -> f16x2;
}

types! {
#![unstable(feature = "stdarch_nvptx", issue = "111199")]

/// PTX-specific 32-bit wide floating point (f16 x 2) vector type
pub struct f16x2(2 x f16);

}

/// Add two values, round to nearest even
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-add>
///
/// Corresponds to the CUDA C intrinsics:
/// - [`__hadd2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g921c795176eaa31265bd80ef4fe4b8e6)
/// - [`__hadd2_rn`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g6cd8ddb2c3d670e1a10c3eb2e7644f82)
#[inline]
#[cfg_attr(test, assert_instr(add.rn.f16x22))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_add(a: f16x2, b: f16x2) -> f16x2 {
simd_add(a, b)
}

/// Subtract two values, round to nearest even
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-sub>
///
/// Corresponds to the CUDA C intrinsics:
/// - [`__hsub2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1ga5536c9c3d853d8c8b9de60e18b41e54)
/// - [`__hsub2_rn`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g8adc164c68d553354f749f0f0645a874)
#[inline]
#[cfg_attr(test, assert_instr(sub.rn.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_sub(a: f16x2, b: f16x2) -> f16x2 {
simd_sub(a, b)
}

/// Multiply two values, round to nearest even
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-mul>
///
/// Corresponds to the CUDA C intrinsics:
/// - [`__hmul2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g70de3f2ee48babe4e0969397ac17708e)
/// - [`__hmul2_rn`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g99f8fe23a4b4c6898d6faf999afaa76e)
#[inline]
#[cfg_attr(test, assert_instr(mul.rn.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_mul(a: f16x2, b: f16x2) -> f16x2 {
simd_mul(a, b)
}

/// Fused multiply-add, round to nearest even
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-fma>
///
/// Corresponds to the CUDA C intrinsics:
/// - [`__fma2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g43628ba21ded8b1e188a367348008dab)
/// - [`__fma2_rn`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__ARITHMETIC.html#group__CUDA__MATH____HALF2__ARITHMETIC_1g43628ba21ded8b1e188a367348008dab)
#[inline]
#[cfg_attr(test, assert_instr(fma.rn.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_fma(a: f16x2, b: f16x2, c: f16x2) -> f16x2 {
simd_fma(a, b, c)
}

/// Arithmetic negate
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-neg>
///
/// Corresponds to the CUDA C intrinsic [`__hmin2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__COMPARISON.html#group__CUDA__MATH____HALF2__COMPARISON_1g9e17a33f96061804166f3fbd395422b6)
#[inline]
#[cfg_attr(test, assert_instr(neg.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_neg(a: f16x2) -> f16x2 {
simd_neg(a)
}

/// Find the minimum of two values
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-min>
///
/// Corresponds to the CUDA C intrinsic [`__hmin2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__COMPARISON.html#group__CUDA__MATH____HALF2__COMPARISON_1g9e17a33f96061804166f3fbd395422b6)
#[inline]
#[cfg_attr(test, assert_instr(min.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_min(a: f16x2, b: f16x2) -> f16x2 {
llvm_f16x2_minnum(a, b)
}

/// Find the minimum of two values, NaNs pass through.
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-min>
///
/// Corresponds to the CUDA C intrinsic [`__hmin2_nan`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__COMPARISON.html#group__CUDA__MATH____HALF2__COMPARISON_1g8bb8f58e9294cc261d2f42c4d5aecd6b)
#[inline]
#[cfg_attr(test, assert_instr(min.NaN.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_min_nan(a: f16x2, b: f16x2) -> f16x2 {
llvm_f16x2_minimum(a, b)
}

/// Find the maximum of two values
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-max>
///
/// Corresponds to the CUDA C intrinsic [`__hmax2`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__COMPARISON.html#group__CUDA__MATH____HALF2__COMPARISON_1g59fc7fc7975d8127b202444a05e57e3d)
#[inline]
#[cfg_attr(test, assert_instr(max.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_max(a: f16x2, b: f16x2) -> f16x2 {
llvm_f16x2_maxnum(a, b)
}

/// Find the maximum of two values, NaNs pass through.
///
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#half-precision-floating-point-instructions-max>
///
/// Corresponds to the CUDA C intrinsic [`__hmax2_nan`](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____HALF2__COMPARISON.html#group__CUDA__MATH____HALF2__COMPARISON_1g41623db7850e3074fd9daa80a14c3897)
#[inline]
#[cfg_attr(test, assert_instr(max.NaN.f16x2))]
#[unstable(feature = "stdarch_nvptx", issue = "111199")]
pub unsafe fn f16x2_max_nan(a: f16x2, b: f16x2) -> f16x2 {
llvm_f16x2_maximum(a, b)
}