arnaucube/poulpy
1
0
Fork 0
mirror of https://github.com/arnaucube/poulpy.git synced 2026-02-10 13:16:44 +01:00
Code Issues Packages Projects Releases Wiki Activity

Ref. + AVX code & generic tests + benches (#85)

Browse Source
This commit is contained in:
Jean-Philippe Bossuat
2025-09-15 16:16:11 +02:00
committed by GitHub
parent 99b9e3e10e
commit 56dbd29c59
286 changed files with 27797 additions and 7270 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
poulpy-hal/src/delegates/vmp_pmat.rs
View File

@@ -1,12 +1,16 @@
use crate::{
api::{
VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftAddTmpBytes, VmpApplyDftToDftTmpBytes, VmpPMatAlloc,
VmpPMatAllocBytes, VmpPMatFromBytes, VmpPrepare, VmpPrepareTmpBytes,
VmpApplyDft, VmpApplyDftTmpBytes, VmpApplyDftToDft, VmpApplyDftToDftAdd, VmpApplyDftToDftAddTmpBytes,
VmpApplyDftToDftTmpBytes, VmpPMatAlloc, VmpPMatAllocBytes, VmpPMatFromBytes, VmpPrepare, VmpPrepareTmpBytes,
},
layouts::{
Backend, MatZnxToRef, Module, Scratch, VecZnxDftToMut, VecZnxDftToRef, VecZnxToRef, VmpPMatOwned, VmpPMatToMut,
VmpPMatToRef,
},
layouts::{Backend, MatZnxToRef, Module, Scratch, VecZnxDftToMut, VecZnxDftToRef, VmpPMatOwned, VmpPMatToMut, VmpPMatToRef},
oep::{
VmpApplyDftToDftAddImpl, VmpApplyDftToDftAddTmpBytesImpl, VmpApplyDftToDftImpl, VmpApplyDftToDftTmpBytesImpl,
VmpPMatAllocBytesImpl, VmpPMatAllocImpl, VmpPMatFromBytesImpl, VmpPMatPrepareImpl, VmpPrepareTmpBytesImpl,
VmpApplyDftImpl, VmpApplyDftTmpBytesImpl, VmpApplyDftToDftAddImpl, VmpApplyDftToDftAddTmpBytesImpl, VmpApplyDftToDftImpl,
VmpApplyDftToDftTmpBytesImpl, VmpPMatAllocBytesImpl, VmpPMatAllocImpl, VmpPMatFromBytesImpl, VmpPrepareImpl,
VmpPrepareTmpBytesImpl,
},
};
@@ -48,7 +52,7 @@ where
impl<B> VmpPrepare<B> for Module<B>
where
B: Backend + VmpPMatPrepareImpl<B>,
B: Backend + VmpPrepareImpl<B>,
{
fn vmp_prepare<R, A>(&self, res: &mut R, a: &A, scratch: &mut Scratch<B>)
where
@@ -59,6 +63,39 @@ where
}
}
impl<B> VmpApplyDftTmpBytes for Module<B>
where
B: Backend + VmpApplyDftTmpBytesImpl<B>,
{
fn vmp_apply_dft_tmp_bytes(
&self,
res_size: usize,
a_size: usize,
b_rows: usize,
b_cols_in: usize,
b_cols_out: usize,
b_size: usize,
) -> usize {
B::vmp_apply_dft_tmp_bytes_impl(
self, res_size, a_size, b_rows, b_cols_in, b_cols_out, b_size,
)
}
}
impl<B> VmpApplyDft<B> for Module<B>
where
B: Backend + VmpApplyDftImpl<B>,
{
fn vmp_apply_dft<R, A, C>(&self, res: &mut R, a: &A, b: &C, scratch: &mut Scratch<B>)
where
R: VecZnxDftToMut<B>,
A: VecZnxToRef,
C: VmpPMatToRef<B>,
{
B::vmp_apply_dft_impl(self, res, a, b, scratch);
}
}
impl<B> VmpApplyDftToDftTmpBytes for Module<B>
where
B: Backend + VmpApplyDftToDftTmpBytesImpl<B>,