floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix avx2 detection 

reword commits to make it clearer
This commit is contained in:
Guillaume Horel 2020-10-31 10:00:48 -04:00
parent 17c46bf06a
commit 1f564d729b
2 changed files with 14 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
cpuid_x86.c
View File

@ -202,7 +202,7 @@ int support_avx(){
if ((ecx & (1 << 28)) != 0 && (ecx & (1 << 27)) != 0 && (ecx & (1 << 26)) != 0){ if ((ecx & (1 << 28)) != 0 && (ecx & (1 << 27)) != 0 && (ecx & (1 << 26)) != 0){
xgetbv(0, &eax, &edx); xgetbv(0, &eax, &edx);
if((eax & 6) == 6){ if((eax & 6) == 6){
ret=1; //OS support AVX ret=1; //OS supports saving xmm and ymm registers (6 = (1<<1) | (1<<2))
} }
} }
return ret; return ret;
@ -219,8 +219,8 @@ int support_avx2(){
if (!support_avx()) if (!support_avx())
return 0; return 0;
cpuid(7, &eax, &ebx, &ecx, &edx); cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & (1<<7)) != 0) if((ebx & (1<<5)) != 0)
ret=1; //OS supports AVX2 ret=1; //CPU supports AVX2
return ret; return ret;
#else #else
return 0; return 0;
@ -235,13 +235,13 @@ int support_avx512(){
if (!support_avx()) if (!support_avx())
return 0; return 0;
cpuid(7, &eax, &ebx, &ecx, &edx); cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & 32) != 32){ if((ebx & (1<<5)) == 0){
ret=0; //OS does not even support AVX2 ret=0; //cpu does not have avx2 flag
} }
if((ebx & (1<<31)) != 0){ if((ebx & (1<<31)) != 0){ //AVX512VL flag
xgetbv(0, &eax, &edx); xgetbv(0, &eax, &edx);
if((eax & 0xe0) == 0xe0) if((eax & 0xe0) == 0xe0)
ret=1; //OS supports AVX512VL ret=1; //OS supports saving zmm registers
} }
return ret; return ret;
#else #else

12
driver/others/dynamic.c
View File

@ -330,8 +330,8 @@ int support_avx2(){
if (!support_avx()) if (!support_avx())
return 0; return 0;
cpuid(7, &eax, &ebx, &ecx, &edx); cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & (1<<7)) != 0) if((ebx & (1<<5)) != 0)
ret=1; //OS supports AVX2 ret=1; //AVX2 flag is set
return ret; return ret;
#else #else
return 0; return 0;
@ -346,13 +346,13 @@ int support_avx512(){
if (!support_avx()) if (!support_avx())
return 0; return 0;
cpuid(7, &eax, &ebx, &ecx, &edx); cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & (1<<7)) == 0){ if((ebx & (1<<5)) == 0){
ret=0; //OS does not even support AVX2 ret=0; //cpu does not have avx2 flag
} }
if((ebx & (1u<<31)) != 0){ if((ebx & (1<<31)) != 0){ //AVX512VL flag is set
xgetbv(0, &eax, &edx); xgetbv(0, &eax, &edx);
if((eax & 0xe0) == 0xe0) if((eax & 0xe0) == 0xe0)
ret=1; //OS supports AVX512VL ret=1; //OS supports saving zmm register
} }
return ret; return ret;
#else #else