Skip to content

arm64: one thread for small level-1 calls on Apple M4 with OpenMP #6254

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
v-code01 wants to merge 1 commit into
base: develop
Choose a base branch
from
+36 −21
Open

arm64: one thread for small level-1 calls on Apple M4 with OpenMP #6254

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 21 additions & 0 deletions common_thread.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,6 +183,27 @@ static __inline int sme_level3_threads(int nthreads) {
#define sme_level3_threads(nthreads) (nthreads)
#endif

#if defined(USE_OPENMP) && defined(ARCH_ARM64) && defined(OS_DARWIN) && (defined(VORTEXM4) || defined(DYNAMIC_ARCH))
#ifdef DYNAMIC_ARCH
extern char *gotoblas_corename(void);
#endif
/* Apple M4 with OpenMP: libomp parks idle threads on Apple Silicon (KMP_BLOCKTIME
defaults to 0 there), and every threaded call wakes the rest of the OpenMP team,
about 3 us per thread on average. A unit-stride level-1 call that reads and writes
fewer than about 384 KB per thread that would be woken is faster on one thread;
bytes is the memory the call reads and writes. Calls with other strides run a
scalar loop and keep the team. Measured on M4 Pro only. */
static __inline int apple_omp_level1_threads(size_t bytes, int unit_stride, int nthreads) {
if (nthreads <= 1 || !unit_stride) return nthreads;
#ifdef DYNAMIC_ARCH
if (strcmp(gotoblas_corename(), "vortexm4") != 0) return nthreads;
#endif
return bytes < (size_t)(nthreads - 1) * 384 * 1024 ? 1 : nthreads;
}
#else
#define apple_omp_level1_threads(bytes, unit_stride, nthreads) (nthreads)
#endif

static __inline int num_cpu_avail(int level) {

#ifdef USE_OPENMP
Expand Down
22 changes: 1 addition & 21 deletions interface/axpy.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,23 +47,6 @@
#define MULTI_THREAD_MINIMAL 10000
#endif

#if defined(SMP) && defined(USE_OPENMP) && defined(ARCH_ARM64) && defined(OS_DARWIN) && \
(defined(DYNAMIC_ARCH) || defined(VORTEXM4))
#ifdef DYNAMIC_ARCH
extern char *gotoblas_corename(void);
#endif
/* Apple M4, OpenMP: libomp parks idle threads on Apple Silicon (KMP_BLOCKTIME
defaults to 0 there), and every call wakes the rest of the OpenMP team, about
3 us per thread on average. Stay on one thread until the vector takes longer
than that, about 128 KB per woken thread; measured on M4 Pro only. */
static int apple_axpy_threads(BLASLONG n, int nthreads) {
#ifdef DYNAMIC_ARCH
if (strcmp(gotoblas_corename(), "vortexm4") != 0) return nthreads;
#endif
return (size_t)n * sizeof(FLOAT) < (size_t)(nthreads - 1) * 128 * 1024 ? 1 : nthreads;
}
#endif

OPENBLAS_EXPORT
#ifndef CBLAS

Expand Down Expand Up @@ -117,10 +100,7 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *y, blasint inc
else
nthreads = num_cpu_avail(1);

#if defined(USE_OPENMP) && defined(ARCH_ARM64) && defined(OS_DARWIN) && \
(defined(DYNAMIC_ARCH) || defined(VORTEXM4))
if (nthreads > 1) nthreads = apple_axpy_threads(n, nthreads);
#endif
nthreads = apple_omp_level1_threads(3 * (size_t)n * sizeof(FLOAT), incx == 1 && incy == 1, nthreads);

if (nthreads == 1) {
#endif
Expand Down
1 change: 1 addition & 0 deletions interface/zaxpy.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,7 @@ void CNAME(blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint in
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads(3 * (size_t)n * COMPSIZE * sizeof(FLOAT), incx == 1 && incy == 1, nthreads);

if (nthreads == 1) {
#endif
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/casum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -237,6 +237,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
asum = casum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/copy_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -188,6 +188,7 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads(2 * (size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1 && inc_y == 1, nthreads);

if (nthreads == 1) {
do_copy(n, x, inc_x, y, inc_y);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/csum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,6 +219,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
asum = casum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/dasum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -234,6 +234,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
asum = dasum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/dot.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,6 +171,7 @@ RETURN_TYPE CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y
nthreads = 1;
else
nthreads = get_dot_optimal_nthreads(n);
nthreads = apple_omp_level1_threads(2 * (size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1 && inc_y == 1, nthreads);

if (nthreads == 1) {
dot = dot_compute(n, x, inc_x, y, inc_y);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/dsum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -216,6 +216,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
dsum = dsum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/iamax_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -325,6 +325,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
max_index = iamax_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/izamax_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -334,6 +334,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
max_index = izamax_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/sasum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,6 +235,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
asum = sasum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/ssum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,6 +217,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
ssum = ssum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/zasum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,6 +235,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
asum = zasum_compute(n, x, inc_x);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/zdot_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -347,6 +347,7 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads(2 * (size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1 && inc_y == 1, nthreads);

if (nthreads == 1) {
zdot_compute(n, x, inc_x, y, inc_y, &zdot);
Expand Down
1 change: 1 addition & 0 deletions kernel/arm64/zsum_thunderx2t99.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,6 +219,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
nthreads = apple_omp_level1_threads((size_t)n * COMPSIZE * sizeof(FLOAT), inc_x == 1, nthreads);

if (nthreads == 1) {
asum = zasum_compute(n, x, inc_x);
Expand Down
Loading