portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

KNL streaming stores, and KNL performance coutners

Browse Source
This commit is contained in:
azusayamaguchi
2016-10-12 11:45:22 +01:00
parent 2d4a45c758
commit 81f2aeaece
12 changed files with 393 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
lib/simd/Intel512common.h
View File

@ -138,9 +138,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define ZLOADf(OFF,PTR,ri,ir) VLOADf(OFF,PTR,ir) VSHUFf(ir,ri)
#define ZLOADd(OFF,PTR,ri,ir) VLOADd(OFF,PTR,ir) VSHUFd(ir,ri)
#define STREAM_STORE
#ifdef STREAM_STORE
#define VSTOREf(OFF,PTR,SRC) "vmovntps " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#define VSTOREd(OFF,PTR,SRC) "vmovntpd " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#else
#define VSTOREf(OFF,PTR,SRC) "vmovaps " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#define VSTOREd(OFF,PTR,SRC) "vmovapd " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#endif
// Swaps Re/Im ; could unify this with IMCI
#define VSHUFd(A,DEST) "vpshufd $0x4e," #A "," #DEST ";\n"