]> git.bitcoin.ninja Git - rust-lightning/commitdiff
Draft NEON that might work
authorMatt Corallo <git@bluematt.me>
Fri, 15 Dec 2023 23:39:57 +0000 (23:39 +0000)
committerMatt Corallo <git@bluematt.me>
Wed, 17 Jul 2024 14:38:21 +0000 (14:38 +0000)
lightning/src/util/simd_f32.rs

index 04d40363ff382d6cda33b36c964435453a7da9fe..fbc3e7951788887deff88b8b0dbfac660b3194b1 100644 (file)
@@ -25,28 +25,28 @@ mod non_simd {
                        (self.3, self.2, self.1, self.0)
                }
        }
-       impl std::ops::Div<FourF32> for FourF32 {
+       impl core::ops::Div<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn div(self, o: FourF32) -> Self {
                        Self(self.0 / o.0, self.1 / o.1, self.2 / o.2, self.3 / o.3)
                }
        }
-       impl std::ops::Mul<FourF32> for FourF32 {
+       impl core::ops::Mul<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn mul(self, o: FourF32) -> Self {
                        Self(self.0 * o.0, self.1 * o.1, self.2 * o.2, self.3 * o.3)
                }
        }
-       impl std::ops::Add<FourF32> for FourF32 {
+       impl core::ops::Add<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn add(self, o: FourF32) -> Self {
                        Self(self.0 + o.0, self.1 + o.1, self.2 + o.2, self.3 + o.3)
                }
        }
-       impl std::ops::Sub<FourF32> for FourF32 {
+       impl core::ops::Sub<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn sub(self, o: FourF32) -> Self {
@@ -60,9 +60,9 @@ pub(crate) use non_simd::*;
 #[cfg(target_feature = "sse")]
 mod x86_sse {
        #[cfg(target_arch = "x86")]
-       use std::arch::x86::*;
+       use core::arch::x86::*;
        #[cfg(target_arch = "x86_64")]
-       use std::arch::x86_64::*;
+       use core::arch::x86_64::*;
 
        #[repr(align(16))]
        struct AlignedFloats([f32; 4]);
@@ -102,28 +102,28 @@ mod x86_sse {
                        (res.0[3], res.0[2], res.0[1], res.0[0])
                }
        }
-       impl std::ops::Div<FourF32> for FourF32 {
+       impl core::ops::Div<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn div(self, o: FourF32) -> Self {
                        Self(unsafe { _mm_div_ps(self.0, o.0) })
                }
        }
-       impl std::ops::Mul<FourF32> for FourF32 {
+       impl core::ops::Mul<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn mul(self, o: FourF32) -> Self {
                        Self(unsafe { _mm_mul_ps(self.0, o.0) })
                }
        }
-       impl std::ops::Add<FourF32> for FourF32 {
+       impl core::ops::Add<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn add(self, o: FourF32) -> Self {
                        Self(unsafe { _mm_add_ps(self.0, o.0) })
                }
        }
-       impl std::ops::Sub<FourF32> for FourF32 {
+       impl core::ops::Sub<FourF32> for FourF32 {
                type Output = FourF32;
                #[inline(always)]
                fn sub(self, o: FourF32) -> Self {
@@ -133,3 +133,76 @@ mod x86_sse {
 }
 #[cfg(target_feature = "sse")]
 pub(crate) use x86_sse::*;
+
+#[cfg(all(target_feature = "neon", target_arch = "aarch64"))]
+mod aarch64_neon {
+       use core::arch::aarch64::*;
+
+       // Not actualy clear if the relevant instructions require alignment, but there's no harm in it
+       // and it may improve performance.
+       #[repr(align(16))]
+       struct AlignedFloats([f32; 4]);
+       #[repr(align(16))]
+       struct AlignedInts([u32; 4]);
+
+       #[derive(Clone, Copy)]
+       pub(crate) struct FourF32(float32x4_t);
+       impl FourF32 {
+               #[inline(always)]
+               pub(crate) fn new(a: f32, b: f32, c: f32, d: f32) -> Self {
+                       let data = AlignedFloats([a, b, c, d]);
+                       Self(unsafe { vld1q_f32(&data.0[0]) })
+               }
+               #[inline(always)]
+               pub(crate) fn from_ints(a: u16, b: u16, c: u16, d: u16) -> Self {
+                       let data = AlignedInts([a as u32, b as u32, c as u32, d as u32]);
+                       let ints = unsafe { vld1q_u32(&data.0[0]) };
+                       Self(unsafe { vcvtq_f32_u32(ints) })
+               }
+               #[inline(always)]
+               pub(crate) fn hsub(&self) -> Self {
+                       let dummy = Self::new(0.0, 0.0, 0.0, 0.0).0; // XXX: There has to be a faster way
+                       Self(unsafe { vpaddq_f32(self.0, dummy) })
+               }
+               #[inline(always)]
+               pub(crate) fn consuming_sum(self) -> f32 {
+                       unsafe { vaddvq_f32(self.0) }
+               }
+               #[inline(always)]
+               pub(crate) fn dump(self) -> (f32, f32, f32, f32) {
+                       let mut res = AlignedFloats([0.0; 4]);
+                       unsafe { vst1q_f32(&mut res.0[0], self.0) };
+                       (res.0[3], res.0[2], res.0[1], res.0[0])
+               }
+       }
+       impl core::ops::Div<FourF32> for FourF32 {
+               type Output = FourF32;
+               #[inline(always)]
+               fn div(self, o: FourF32) -> Self {
+                       Self(unsafe { vdivq_f32(self.0, o.0) })
+               }
+       }
+       impl core::ops::Mul<FourF32> for FourF32 {
+               type Output = FourF32;
+               #[inline(always)]
+               fn mul(self, o: FourF32) -> Self {
+                       Self(unsafe { vmulq_f32(self.0, o.0) })
+               }
+       }
+       impl core::ops::Add<FourF32> for FourF32 {
+               type Output = FourF32;
+               #[inline(always)]
+               fn add(self, o: FourF32) -> Self {
+                       Self(unsafe { vaddq_f32(self.0, o.0) })
+               }
+       }
+       impl core::ops::Sub<FourF32> for FourF32 {
+               type Output = FourF32;
+               #[inline(always)]
+               fn sub(self, o: FourF32) -> Self {
+                       Self(unsafe { vsubq_f32(self.0, o.0) })
+               }
+       }
+}
+#[cfg(all(target_feature = "neon", target_arch = "aarch64"))]
+pub(crate) use aarch64_neon::*;