Update to custom fheuint prepare

poulpy-hal/src/api/vmp_pmat.rs

@@ -111,3 +111,9 @@ pub trait VmpApplyDftToDftAdd<B: Backend> {
         A: VecZnxDftToRef<B>,
         C: VmpPMatToRef<B>;
 }
+
+pub trait VmpZero<B: Backend> {
+    fn vmp_zero<R>(&self, res: &mut R)
+    where
+        R: VmpPMatToMut<B>;
+}

poulpy-hal/src/delegates/vmp_pmat.rs

@@ -1,7 +1,7 @@
 use crate::{
     api::{
         VmpApplyDft, VmpApplyDftTmpBytes, VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftAddTmpBytes,
-        VmpApplyDftToDftTmpBytes, VmpPMatAlloc, VmpPMatBytesOf, VmpPMatFromBytes, VmpPrepare, VmpPrepareTmpBytes,
+        VmpApplyDftToDftTmpBytes, VmpPMatAlloc, VmpPMatBytesOf, VmpPMatFromBytes, VmpPrepare, VmpPrepareTmpBytes, VmpZero,
     },
     layouts::{
         Backend, MatZnxToRef, Module, Scratch, VecZnxDftToMut, VecZnxDftToRef, VecZnxToRef, VmpPMatOwned, VmpPMatToMut,
@@ -10,7 +10,7 @@ use crate::{
     oep::{
         VmpApplyDftImpl, VmpApplyDftTmpBytesImpl, VmpApplyDftToDftAddImpl, VmpApplyDftToDftAddTmpBytesImpl, VmpApplyDftToDftImpl,
         VmpApplyDftToDftTmpBytesImpl, VmpPMatAllocBytesImpl, VmpPMatAllocImpl, VmpPMatFromBytesImpl, VmpPrepareImpl,
-        VmpPrepareTmpBytesImpl,
+        VmpPrepareTmpBytesImpl, VmpZeroImpl,
     },
 };
 
@@ -161,3 +161,15 @@ where
         B::vmp_apply_dft_to_dft_add_impl(self, res, a, b, scale, scratch);
     }
 }
+
+impl<B> VmpZero<B> for Module<B>
+where
+    B: Backend + VmpZeroImpl<B>,
+{
+    fn vmp_zero<R>(&self, res: &mut R)
+    where
+        R: VmpPMatToMut<B>,
+    {
+        B::vmp_zero_impl(self, res);
+    }
+}

poulpy-hal/src/oep/vmp_pmat.rs

@@ -145,3 +145,13 @@ pub unsafe trait VmpApplyDftToDftAddImpl<B: Backend> {
         A: VecZnxDftToRef<B>,
         C: VmpPMatToRef<B>;
 }
+
+/// # THIS TRAIT IS AN OPEN EXTENSION POINT (unsafe)
+/// * See TODO.
+/// * See [crate::api::VmpZero] for corresponding public API.
+/// # Safety [crate::doc::backend_safety] for safety contract.
+pub unsafe trait VmpZeroImpl<B: Backend> {
+    fn vmp_zero_impl<R>(module: &Module<B>, res: &mut R)
+    where
+        R: VmpPMatToMut<B>;
+}

poulpy-hal/src/reference/fft64/vmp.rs

@@ -1,6 +1,6 @@
 use crate::{
     cast_mut,
-    layouts::{MatZnx, MatZnxToRef, VecZnx, VecZnxToRef, VmpPMatToMut, ZnxView, ZnxViewMut},
+    layouts::{DataViewMut, MatZnx, MatZnxToRef, VecZnx, VecZnxToRef, VmpPMatToMut, ZnxView, ZnxViewMut},
     oep::VecZnxDftAllocBytesImpl,
     reference::fft64::{
         reim::{ReimDFTExecute, ReimFFTTable, ReimFromZnx, ReimZero},
@@ -157,6 +157,13 @@ pub fn vmp_apply_dft_to_dft_tmp_bytes(a_size: usize, prows: usize, pcols_in: usi
     (16 + 8 * row_max * pcols_in) * size_of::<f64>()
 }
 
+pub fn vmp_zero<R, BE: Backend>(res: &mut R)
+where
+    R: VmpPMatToMut<BE>,
+{
+    res.to_mut().data_mut().fill(0);
+}
+
 pub fn vmp_apply_dft_to_dft<R, A, M, BE>(res: &mut R, a: &A, pmat: &M, tmp_bytes: &mut [f64])
 where
     BE: Backend<ScalarPrep = f64>