full piv lu

sarah-quinones · Oct 12, 2024 · bbdbb44 · bbdbb44
1 parent 1ee3003
commit bbdbb44
Show file tree

Hide file tree

Showing 4 changed files with 904 additions and 7 deletions.
diff --git a/faer-traits/src/lib.rs b/faer-traits/src/lib.rs
@@ -340,6 +340,34 @@ macro_rules! help2 {
             }};
         }
 
+        #[allow(unused_macros)]
+        macro_rules! send2 {
+            ($place: expr) => {
+                $crate::utils::send::<$C, _>($place)
+            };
+        }
+
+        #[allow(unused_macros)]
+        macro_rules! unsend2 {
+            ($place: expr) => {
+                $crate::utils::unsend::<$C, _>($place)
+            };
+        }
+
+        #[allow(unused_macros)]
+        macro_rules! sync2 {
+            ($place: expr) => {
+                $crate::utils::sync::<$C, _>($place)
+            };
+        }
+
+        #[allow(unused_macros)]
+        macro_rules! unsync2 {
+            ($place: expr) => {
+                $crate::utils::unsync::<$C, _>($place)
+            };
+        }
+
         #[allow(unused_macros)]
         macro_rules! by_ref2 {
             ($place: expr) => {
@@ -1072,6 +1100,10 @@ impl<C: ComplexContainer, T: ComplexField<C>, S: Simd> SimdCtx<C, T, S> {
     pub fn and_mask(&self, lhs: T::SimdMask<S>, rhs: T::SimdMask<S>) -> T::SimdMask<S> {
         T::simd_and_mask(&self.0, lhs, rhs)
     }
+    #[inline(always)]
+    pub fn first_true_mask(&self, value: T::SimdMask<S>) -> usize {
+        T::simd_first_true_mask(&self.0, value)
+    }
 
     #[inline(always)]
     pub unsafe fn mask_load(
@@ -1365,6 +1397,10 @@ impl<C: ComplexContainer, T: ComplexField<C>, S: Simd> SimdCtxCopy<C, T, S> {
         T::simd_and_mask(&self.0, lhs, rhs)
     }
     #[inline(always)]
+    pub fn first_true_mask(&self, value: T::SimdMask<S>) -> usize {
+        T::simd_first_true_mask(&self.0, value)
+    }
+    #[inline(always)]
     pub unsafe fn mask_load(
         &self,
         mask: T::SimdMask<S>,
@@ -2290,6 +2326,8 @@ pub trait ComplexField<C: ComplexContainer = Unit>:
         lhs: Self::SimdMask<S>,
         rhs: Self::SimdMask<S>,
     ) -> Self::SimdMask<S>;
+    fn simd_first_true_mask<S: Simd>(ctx: &Self::SimdCtx<S>, value: Self::SimdMask<S>) -> usize;
+
     fn simd_head_mask<S: Simd>(ctx: &Self::SimdCtx<S>, len: usize) -> Self::SimdMask<S>;
     unsafe fn simd_mask_load<S: Simd>(
         ctx: &Self::SimdCtx<S>,
@@ -3037,6 +3075,11 @@ impl<C: RealContainer, T: RealField<C>> ComplexField<Complex<C>> for T {
     ) -> Self::SimdMask<S> {
         T::simd_and_mask(ctx, lhs, rhs)
     }
+
+    #[inline(always)]
+    fn simd_first_true_mask<S: Simd>(ctx: &Self::SimdCtx<S>, value: Self::SimdMask<S>) -> usize {
+        T::simd_first_true_mask(ctx, value)
+    }
 }
 
 impl AsRef<Unit> for Unit {
@@ -3453,6 +3496,11 @@ impl ComplexField for f32 {
     ) -> Self::SimdMask<S> {
         simd.and_m32s(lhs, rhs)
     }
+
+    #[inline(always)]
+    fn simd_first_true_mask<S: Simd>(ctx: &Self::SimdCtx<S>, value: Self::SimdMask<S>) -> usize {
+        ctx.first_true_m32s(value)
+    }
 }
 
 impl RealField for f32 {
@@ -3891,6 +3939,11 @@ impl ComplexField for f64 {
     ) -> Self::SimdMask<S> {
         simd.and_m64s(lhs, rhs)
     }
+
+    #[inline(always)]
+    fn simd_first_true_mask<S: Simd>(ctx: &Self::SimdCtx<S>, value: Self::SimdMask<S>) -> usize {
+        ctx.first_true_m64s(value)
+    }
 }
 
 impl RealField for f64 {
@@ -4405,6 +4458,11 @@ impl<T: EnableComplex> ComplexField for Complex<T> {
     ) -> Self::SimdMask<S> {
         T::simd_and_mask(simd, lhs, rhs)
     }
+
+    #[inline(always)]
+    fn simd_first_true_mask<S: Simd>(ctx: &Self::SimdCtx<S>, value: Self::SimdMask<S>) -> usize {
+        T::simd_first_true_mask(ctx, value)
+    }
 }
 
 impl EnableComplex for f32 {

diff --git a/faer/src/linalg/cholesky/ldlt/factor.rs b/faer/src/linalg/cholesky/ldlt/factor.rs
@@ -497,7 +497,7 @@ pub(crate) fn cholesky_recursion<'N, C: ComplexContainer, T: ComplexField<C>>(
             ghost_tree!(FULL(MAT(HEAD, TAIL)), {
                 let (full, FULL) = N.full(FULL);
                 let j = full.from_local(j);
-                let (mat, MAT) = full.segment(j.into(), full.from_local_inc(j_next), FULL.MAT);
+                let (mat, MAT) = full.segment(j.into(), full.from_local_inc(N.end()), FULL.MAT);
 
                 let j_next = mat.idx_inc(*j_next);
                 let (disjoint, head, tail, _, _) = mat.split_inc(j_next, MAT.HEAD, MAT.TAIL);
@@ -637,8 +637,8 @@ impl Default for LdltParams {
     #[inline]
     fn default() -> Self {
         Self {
-            recursion_threshold: NonZero::new(64).unwrap(),
-            blocksize: NonZero::new(128).unwrap(),
+            recursion_threshold: NonZero::new(2).unwrap(),
+            blocksize: NonZero::new(2).unwrap(),
         }
     }
 }
@@ -776,7 +776,7 @@ mod tests {
         type C = faer_traits::Unit;
         type T = c64;
 
-        for n in [2, 127, 240] {
+        for n in [2, 4, 8, 31, 127, 240] {
             with_dim!(N, n);
 
             for llt in [false, true] {