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

Merge pull request #1846 from fenrus75/threadsize 

gemm/dgemm: add a way for an arch kernel to specify preferred sizes
This commit is contained in:
Martin Kroeker 2018-11-02 13:18:01 +01:00 committed by GitHub
parent 661035477c 5b708e5eb1
commit f1c02273cb
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 31 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
driver/level3/level3_thread.c
View File

@ -48,6 +48,10 @@
#define SWITCH_RATIO 2
#endif
#ifndef GEMM_PREFERED_SIZE
#define GEMM_PREFERED_SIZE 1
#endif
//The array of job_t may overflow the stack.
//Instead, use malloc to alloc job_t.
#if MAX_CPU_NUMBER > BLAS3_MEM_ALLOC_THRESHOLD
@ -510,6 +514,16 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
return 0;
}
static int round_up(int remainder, int width, int multiple)
{
if (multiple > remainder || width <= multiple)
return width;
width = (width + multiple - 1) / multiple;
width = width * multiple;
return width;
}
static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
*range_n, FLOAT *sa, FLOAT *sb,
BLASLONG nthreads_m, BLASLONG nthreads_n) {
@ -601,9 +615,14 @@ static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
num_parts = 0;
while (m > 0){
width = blas_quickdivide(m + nthreads_m - num_parts - 1, nthreads_m - num_parts);
width = round_up(m, width, GEMM_PREFERED_SIZE);
m -= width;
if (m < 0) width = width + m;
range_M[num_parts + 1] = range_M[num_parts] + width;
num_parts ++;
}
for (i = num_parts; i < MAX_CPU_NUMBER; i++) {
@ -645,9 +664,12 @@ static int gemm_driver(blas_arg_t *args, BLASLONG *range_m, BLASLONG
if (width < SWITCH_RATIO) {
width = SWITCH_RATIO;
}
width = round_up(n, width, GEMM_PREFERED_SIZE);
n -= width;
if (n < 0) width = width + n;
range_N[num_parts + 1] = range_N[num_parts] + width;
num_parts ++;
}
for (j = num_parts; j < MAX_CPU_NUMBER; j++) {

6
kernel/x86_64/dgemm_beta_skylakex.c
View File

@ -55,6 +55,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
return 0;
}
if (m == 0 || n == 0)
return 0;
c_offset = c;
@ -69,7 +71,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
i = m;
while (i > 32) {
while (i >= 32) {
_mm512_storeu_pd(c_offset1, z_zero);
_mm512_storeu_pd(c_offset1 + 8, z_zero);
_mm512_storeu_pd(c_offset1 + 16, z_zero);
@ -77,7 +79,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
c_offset1 += 32;
i -= 32;
}
while (i > 8) {
while (i >= 8) {
_mm512_storeu_pd(c_offset1, z_zero);
c_offset1 += 8;
i -= 8;

6
kernel/x86_64/sgemm_beta_skylakex.c
View File

@ -55,6 +55,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
return 0;
}
if (n == 0 || m == 0)
return;
c_offset = c;
@ -71,13 +73,13 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
i = m;
while (i > 32) {
while (i >= 32) {
_mm512_storeu_ps(c_offset1, z_zero);
_mm512_storeu_ps(c_offset1 + 16, z_zero);
c_offset1 += 32;
i -= 32;
}
while (i > 8) {
while (i >= 8) {
_mm256_storeu_ps(c_offset1, y_zero);
c_offset1 += 8;
i -= 8;

1
param.h
View File

@ -1627,6 +1627,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define SYMV_P 8
#define SWITCH_RATIO 32
#define GEMM_PREFERED_SIZE 32
#ifdef ARCH_X86