From f926ace25bfba1ed3488f3de60b36f5a47d95fc9 Mon Sep 17 00:00:00 2001 From: Martin Reinecke Date: Wed, 31 Jul 2019 15:12:22 +0200 Subject: [PATCH] more minor tweaks --- pocketfft_hdronly.h | 43 ++++++++++++++++++------------------------- 1 file changed, 18 insertions(+), 25 deletions(-) diff --git a/pocketfft_hdronly.h b/pocketfft_hdronly.h index 172cc54..0c90f55 100644 --- a/pocketfft_hdronly.h +++ b/pocketfft_hdronly.h @@ -926,15 +926,14 @@ template void pass8 (size_t ido, size_t l1, PMC(a0,a4,CC(0,0,k),CC(0,4,k)); PMC(a2,a6,CC(0,2,k),CC(0,6,k)); ROTX90(a6); - PMINPLACE(a0,a2); - PMINPLACE(a4,a6); - PMC(CH(0,k,0),CH(0,k,4),a0,a1); - PMC(CH(0,k,1),CH(0,k,5),a4,a5); - PMC(CH(0,k,2),CH(0,k,6),a2,a3); - PMC(CH(0,k,3),CH(0,k,7),a6,a7); + PMC(CH(0,k,0),CH(0,k,4),a0+a2,a1); + PMC(CH(0,k,1),CH(0,k,5),a4+a6,a5); + PMC(CH(0,k,2),CH(0,k,6),a0-a2,a3); + PMC(CH(0,k,3),CH(0,k,7),a4-a6,a7); } else for (size_t k=0; k void pass8 (size_t ido, size_t l1, PMC(a0,a4,CC(0,0,k),CC(0,4,k)); PMC(a2,a6,CC(0,2,k),CC(0,6,k)); ROTX90(a6); - PMINPLACE(a0,a2); - PMINPLACE(a4,a6); - PMC(CH(0,k,0),CH(0,k,4),a0,a1); - PMC(CH(0,k,1),CH(0,k,5),a4,a5); - PMC(CH(0,k,2),CH(0,k,6),a2,a3); - PMC(CH(0,k,3),CH(0,k,7),a6,a7); - + PMC(CH(0,k,0),CH(0,k,4),a0+a2,a1); + PMC(CH(0,k,1),CH(0,k,5),a4+a6,a5); + PMC(CH(0,k,2),CH(0,k,6),a0-a2,a3); + PMC(CH(0,k,3),CH(0,k,7),a4-a6,a7); + } for (size_t i=1; i void pass8 (size_t ido, size_t l1, ROTX90(a6); PMINPLACE(a0,a2); PMINPLACE(a4,a6); - PMINPLACE(a0,a1); - CH(i,k,0) = a0; - CH(i,k,4) = a1.template special_mul(WA(3,i)); - PMINPLACE(a2,a3); - CH(i,k,2) = a2.template special_mul(WA(1,i)); - CH(i,k,6) = a3.template special_mul(WA(5,i)); - PMINPLACE(a4,a5); - CH(i,k,1) = a4.template special_mul(WA(0,i)); - CH(i,k,5) = a5.template special_mul(WA(4,i)); - PMINPLACE(a6,a7); - CH(i,k,3) = a6.template special_mul(WA(2,i)); - CH(i,k,7) = a7.template special_mul(WA(6,i)); + CH(i,k,0) = a0+a1; + CH(i,k,4) = (a0-a1).template special_mul(WA(3,i)); + CH(i,k,1) = (a4+a5).template special_mul(WA(0,i)); + CH(i,k,5) = (a4-a5).template special_mul(WA(4,i)); + CH(i,k,2) = (a2+a3).template special_mul(WA(1,i)); + CH(i,k,6) = (a2-a3).template special_mul(WA(5,i)); + CH(i,k,3) = (a6+a7).template special_mul(WA(2,i)); + CH(i,k,7) = (a6-a7).template special_mul(WA(6,i)); } } } -- GitLab