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

Merge pull request #1949 from martin-frbg/issue1947 

Query AVX2 and AVX512VL support when selecting x86 kernels
This commit is contained in:
Martin Kroeker 2019-01-08 20:44:08 +01:00 committed by GitHub
parent cf5d48e833 1650311246
commit 8d99dba86b
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 186 additions and 98 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
common_x86_64.h
View File

@ -134,7 +134,7 @@ static __inline void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
"=b" (*ebx),
"=c" (*ecx),
"=d" (*edx)
: "0" (op));
: "0" (op), "c"(0));
#endif
}

1
cpuid.h
View File

@ -139,6 +139,7 @@
#define HAVE_FMA4 (1 << 19)
#define HAVE_FMA3 (1 << 20)
#define HAVE_AVX512VL (1 << 21)
#define HAVE_AVX2 (1 << 22)
#define CACHE_INFO_L1_I 1
#define CACHE_INFO_L1_D 2

134
cpuid_x86.c
View File

@ -97,10 +97,10 @@ static C_INLINE void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
("mov %%ebx, %%edi;"
"cpuid;"
"xchgl %%ebx, %%edi;"
: "=a" (*eax), "=D" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) : "cc");
: "=a" (*eax), "=D" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op), "c" (0) : "cc");
#else
__asm__ __volatile__
("cpuid": "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) : "cc");
("cpuid": "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) , "c" (0) : "cc");
#endif
}
@ -211,6 +211,44 @@ int support_avx(){
#endif
}
int support_avx2(){
#ifndef NO_AVX2
int eax, ebx, ecx=0, edx;
int ret=0;
if (!support_avx)
return 0;
cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & (1<<7)) != 0)
ret=1; //OS supports AVX2
return ret;
#else
return 0;
#endif
}
int support_avx512(){
#ifndef NO_AVX512
int eax, ebx, ecx, edx;
int ret=0;
if (!support_avx)
return 0;
cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & 32) != 32){
ret=0; //OS does not even support AVX2
}
if((ebx & (1<<31)) != 0){
xgetbv(0, &eax, &edx);
if((eax & 0xe0) == 0xe0)
ret=1; //OS supports AVX512VL
}
return ret;
#else
return 0;
#endif
}
int get_vendor(void){
int eax, ebx, ecx, edx;
@ -294,6 +332,8 @@ int get_cputype(int gettype){
if ((ecx & (1 << 20)) != 0) feature |= HAVE_SSE4_2;
#ifndef NO_AVX
if (support_avx()) feature |= HAVE_AVX;
if (support_avx2()) feature |= HAVE_AVX2;
if (support_avx512()) feature |= HAVE_AVX512VL;
if ((ecx & (1 << 12)) != 0) feature |= HAVE_FMA3;
#endif
@ -1228,22 +1268,18 @@ int get_cpuname(void){
return CPUTYPE_NEHALEM;
case 12:
case 15:
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 13:
//Broadwell
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
}
@ -1252,33 +1288,27 @@ int get_cpuname(void){
switch (model) {
case 5:
case 6:
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 7:
case 15:
//Broadwell
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 14:
//Skylake
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 12:
@ -1292,46 +1322,36 @@ int get_cpuname(void){
switch (model) {
case 6:
//Broadwell
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 5:
// Skylake X
#ifndef NO_AVX512
return CPUTYPE_SKYLAKEX;
#else
if(support_avx())
#ifndef NO_AVX2
return CPUTYPE_HASWELL;
#else
return CPUTYPE_SANDYBRIDGE;
#endif
if(support_avx512())
return CPUTYPE_SKYLAKEX;
if(support_avx2())
return CPUTYPE_HASWELL;
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
else
return CPUTYPE_NEHALEM;
#endif
case 14:
// Skylake
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 7:
// Xeon Phi Knights Landing
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
case 12:
@ -1342,30 +1362,24 @@ int get_cpuname(void){
case 6:
switch (model) {
case 6: // Cannon Lake
#ifndef NO_AVX512
return CPUTYPE_SKYLAKEX;
#else
if(support_avx())
#ifndef NO_AVX2
return CPUTYPE_HASWELL;
#else
return CPUTYPE_SANDYBRIDGE;
#endif
if(support_avx512())
return CPUTYPE_SKYLAKEX;
if(support_avx2())
return CPUTYPE_HASWELL;
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
else
return CPUTYPE_NEHALEM;
#endif
}
break;
case 9:
case 8:
switch (model) {
case 14: // Kaby Lake
if(support_avx())
#ifndef NO_AVX2
if(support_avx2())
return CPUTYPE_HASWELL;
#else
if(support_avx())
return CPUTYPE_SANDYBRIDGE;
#endif
else
return CPUTYPE_NEHALEM;
}
@ -2112,6 +2126,8 @@ void get_cpuconfig(void){
if (features & HAVE_SSE4A) printf("#define HAVE_SSE4A\n");
if (features & HAVE_SSE5 ) printf("#define HAVE_SSSE5\n");
if (features & HAVE_AVX ) printf("#define HAVE_AVX\n");
if (features & HAVE_AVX2 ) printf("#define HAVE_AVX2\n");
if (features & HAVE_AVX512VL ) printf("#define HAVE_AVX512VL\n");
if (features & HAVE_3DNOWEX) printf("#define HAVE_3DNOWEX\n");
if (features & HAVE_3DNOW) printf("#define HAVE_3DNOW\n");
if (features & HAVE_FMA4 ) printf("#define HAVE_FMA4\n");
@ -2180,6 +2196,8 @@ void get_sse(void){
if (features & HAVE_SSE4A) printf("HAVE_SSE4A=1\n");
if (features & HAVE_SSE5 ) printf("HAVE_SSSE5=1\n");
if (features & HAVE_AVX ) printf("HAVE_AVX=1\n");
if (features & HAVE_AVX2 ) printf("HAVE_AVX2=1\n");
if (features & HAVE_AVX512VL ) printf("HAVE_AVX512VL=1\n");
if (features & HAVE_3DNOWEX) printf("HAVE_3DNOWEX=1\n");
if (features & HAVE_3DNOW) printf("HAVE_3DNOW=1\n");
if (features & HAVE_FMA4 ) printf("HAVE_FMA4=1\n");

147
driver/others/dynamic.c
View File

@ -304,9 +304,49 @@ int support_avx(){
#endif
}
int support_avx2(){
#ifndef NO_AVX2
int eax, ebx, ecx=0, edx;
int ret=0;
if (!support_avx)
return 0;
cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & (1<<7)) != 0)
ret=1; //OS supports AVX2
return ret;
#else
return 0;
#endif
}
int support_avx512(){
#ifndef NO_AVX512
int eax, ebx, ecx, edx;
int ret=0;
if (!support_avx)
return 0;
cpuid(7, &eax, &ebx, &ecx, &edx);
if((ebx & (1<<7)) != 1){
ret=0; //OS does not even support AVX2
}
if((ebx & (1<<31)) != 0){
xgetbv(0, &eax, &edx);
if((eax & 0xe0) == 0xe0)
ret=1; //OS supports AVX512VL
}
return ret;
#else
return 0;
#endif
}
extern void openblas_warning(int verbose, const char * msg);
#define FALLBACK_VERBOSE 1
#define NEHALEM_FALLBACK "OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Nehalem kernels as a fallback, which may give poorer performance.\n"
#define SANDYBRIDGE_FALLBACK "OpenBLAS : Your OS does not support AVX2 instructions. OpenBLAS is using Sandybridge kernels as a fallback, which may give poorer performance.\n"
#define HASWELL_FALLBACK "OpenBLAS : Your OS does not support AVX512VL instructions. OpenBLAS is using Haswell kernels as a fallback, which may give poorer performance.\n"
#define BARCELONA_FALLBACK "OpenBLAS : Your OS does not support AVX instructions. OpenBLAS is using Barcelona kernels as a fallback, which may give poorer performance.\n"
static int get_vendor(void){
@ -403,18 +443,24 @@ static gotoblas_t *get_coretype(void){
}
//Intel Haswell
if (model == 12 || model == 15) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
}
//Intel Broadwell
if (model == 13) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
@ -424,27 +470,36 @@ static gotoblas_t *get_coretype(void){
case 4:
//Intel Haswell
if (model == 5 || model == 6) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
}
//Intel Broadwell
if (model == 7 || model == 15) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
}
//Intel Skylake
if (model == 14) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
@ -457,40 +512,54 @@ static gotoblas_t *get_coretype(void){
case 5:
//Intel Broadwell
if (model == 6) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
}
if (model == 5) {
// Intel Skylake X
#ifndef NO_AVX512
return &gotoblas_SKYLAKEX;
#else
if(support_avx())
if (support_avx512())
return &gotoblas_SKYLAKEX;
if(support_avx2()){
openblas_warning(FALLBACK_VERBOSE, HASWELL_FALLBACK);
return &gotoblas_HASWELL;
else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM;
}
#endif
}
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM;
}
}
//Intel Skylake
if (model == 14) {
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
}
//Intel Phi Knights Landing
if (model == 7) {
if(support_avx())
if(support_avx2()){
openblas_warning(FALLBACK_VERBOSE, HASWELL_FALLBACK);
return &gotoblas_HASWELL;
else{
}
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}
@ -503,26 +572,26 @@ static gotoblas_t *get_coretype(void){
case 6:
if (model == 6) {
// Cannon Lake
#ifndef NO_AVX512
return &gotoblas_SKYLAKEX;
#else
if(support_avx())
#ifndef NO_AVX2
return &gotoblas_HASWELL;
#else
return &gotoblas_SANDYBRIDGE;
#endif
else
return &gotoblas_NEHALEM;
#endif
if(support_avx2())
return &gotoblas_HASWELL;
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM;
}
}
return NULL;
case 9:
case 8:
if (model == 14 ) { // Kaby Lake
if(support_avx())
if(support_avx2())
return &gotoblas_HASWELL;
else{
if(support_avx()) {
openblas_warning(FALLBACK_VERBOSE, SANDYBRIDGE_FALLBACK);
return &gotoblas_SANDYBRIDGE;
} else {
openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK);
return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
}