Skip to content

Multi-thread Performance Improvement of GEMM on NeoverseV1 with DIVIDE_RATE=1 #5407

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

Merged
martin-frbg merged 1 commit into from
Jul 30, 2025
Merged

Multi-thread Performance Improvement of GEMM on NeoverseV1 with DIVIDE_RATE=1 #5407

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
4 changes: 4 additions & 0 deletions driver/level3/gemm.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,6 +63,10 @@
#define DIVIDE_RATE GEMM_DIVIDE_RATE
#endif

#ifdef GEMM_DIVIDE_LIMIT
#define DIVIDE_LIMIT GEMM_DIVIDE_LIMIT
#endif

#ifdef THREADED_LEVEL3
#include "level3_thread.c"
#else
Expand Down
18 changes: 12 additions & 6 deletions driver/level3/level3_thread.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,6 +246,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,

BLASLONG nthreads_m;
BLASLONG mypos_m, mypos_n;
BLASLONG divide_rate = DIVIDE_RATE;

BLASLONG is, js, ls, bufferside, jjs;
BLASLONG min_i, min_l, div_n, min_jj;
Expand Down Expand Up @@ -280,6 +281,11 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
alpha = (FLOAT *)args -> alpha;
beta = (FLOAT *)args -> beta;

/* Disable divide_rate when N of all threads are less than to DIVIDE_LIMIT */
#ifdef DIVIDE_LIMIT
if (N < DIVIDE_LIMIT) divide_rate = 1;
#endif

/* Initialize 2D CPU distribution */
nthreads_m = args -> nthreads;
if (range_m) {
Expand Down Expand Up @@ -321,9 +327,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
) return 0;

/* Initialize workspace for local region of B */
div_n = (n_to - n_from + DIVIDE_RATE - 1) / DIVIDE_RATE;
div_n = (n_to - n_from + divide_rate - 1) / divide_rate;
buffer[0] = sb;
for (i = 1; i < DIVIDE_RATE; i++) {
for (i = 1; i < divide_rate; i++) {
buffer[i] = buffer[i - 1] + GEMM_Q * ((div_n + GEMM_UNROLL_N - 1)/GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE;
}

Expand Down Expand Up @@ -365,7 +371,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
STOP_RPCC(copy_A);

/* Copy local region of B into workspace and apply kernel */
div_n = (n_to - n_from + DIVIDE_RATE - 1) / DIVIDE_RATE;
div_n = (n_to - n_from + divide_rate - 1) / divide_rate;
for (js = n_from, bufferside = 0; js < n_to; js += div_n, bufferside ++) {

/* Make sure if no one is using workspace */
Expand Down Expand Up @@ -434,7 +440,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
if (current >= (mypos_n + 1) * nthreads_m) current = mypos_n * nthreads_m;

/* Split other region of B into parts */
div_n = (range_n[current + 1] - range_n[current] + DIVIDE_RATE - 1) / DIVIDE_RATE;
div_n = (range_n[current + 1] - range_n[current] + divide_rate - 1) / divide_rate;
for (js = range_n[current], bufferside = 0; js < range_n[current + 1]; js += div_n, bufferside ++) {
if (current != mypos) {

Expand Down Expand Up @@ -485,7 +491,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
do {

/* Split region of B into parts and apply kernel */
div_n = (range_n[current + 1] - range_n[current] + DIVIDE_RATE - 1) / DIVIDE_RATE;
div_n = (range_n[current + 1] - range_n[current] + divide_rate - 1) / divide_rate;
for (js = range_n[current], bufferside = 0; js < range_n[current + 1]; js += div_n, bufferside ++) {

/* Apply kernel with local region of A and part of region of B */
Expand Down Expand Up @@ -520,7 +526,7 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
/* Wait until all other threads are done with local region of B */
START_RPCC();
for (i = 0; i < args -> nthreads; i++) {
for (js = 0; js < DIVIDE_RATE; js++) {
for (js = 0; js < divide_rate; js++) {
while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;};
}
}
Expand Down
2 changes: 2 additions & 0 deletions param.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3585,6 +3585,8 @@ is a big desktop or server with abundant cache rather than a phone or embedded d

#elif defined(NEOVERSEV1) // 256-bit SVE

#define GEMM_DIVIDE_LIMIT 3

#if defined(XDOUBLE) || defined(DOUBLE)
#define SWITCH_RATIO 8
#define GEMM_PREFERED_SIZE 4
Expand Down
Loading