Skip to content

Dgemm loop unroll and 4x1, 4x2 dgemv VSX implementation for power10. #5453

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

Draft
wants to merge 2 commits into
base: develop
Choose a base branch
from
+107 −16
Draft

Dgemm loop unroll and 4x1, 4x2 dgemv VSX implementation for power10. #5453

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
16be28a
dgemm loop unroll and 4x1 4x2 dgemv implimentation
pratiklp00 Sep 22, 2025
e2399be
add macro
pratiklp00 Oct 9, 2025
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
78 changes: 62 additions & 16 deletions kernel/power/dgemm_kernel_power10.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,7 +92,18 @@ typedef FLOAT v4sf_t __attribute__ ((vector_size (16)));
rowC = (v4sf_t *) &CO[1* ldc+J]; \
rowC[0] += result[1] * alpha;
#endif

#define KERNEL(i) \
rowA = (vec_t *)&AO[i<< 3];\
rowB = *((__vector_pair *)((void *)&BO[i << 3]));\
rowB1 = *((__vector_pair *)((void *)&BO[(i << 3) + 4]));\
__builtin_mma_xvf64gerpp(&acc0, rowB, rowA[0]);\
__builtin_mma_xvf64gerpp(&acc1, rowB1, rowA[0]);\
__builtin_mma_xvf64gerpp(&acc2, rowB, rowA[1]);\
__builtin_mma_xvf64gerpp(&acc3, rowB1, rowA[1]);\
__builtin_mma_xvf64gerpp(&acc4, rowB, rowA[2]);\
__builtin_mma_xvf64gerpp(&acc5, rowB1, rowA[2]);\
__builtin_mma_xvf64gerpp(&acc6, rowB, rowA[3]);\
__builtin_mma_xvf64gerpp(&acc7, rowB1, rowA[3]);
#define PREFETCH1(x, y) asm volatile ("dcbt %0, %1" : : "r" (x), "b" (y) : "memory");

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
Expand Down Expand Up @@ -188,7 +199,7 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
v4sf_t *rowC;
v4sf_t result[4];
__vector_quad acc0, acc1, acc2, acc3, acc4,acc5,acc6,acc7;
BLASLONG l = 0;
BLASLONG l = 1;
vec_t *rowA = (vec_t *) & AO[0];
__vector_pair rowB, rowB1;
rowB = *((__vector_pair *)((void *)&BO[0]));
Expand All @@ -201,20 +212,55 @@ CNAME (BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT * A, FLOAT * B,
__builtin_mma_xvf64ger (&acc5, rowB1, rowA[2]);
__builtin_mma_xvf64ger (&acc6, rowB, rowA[3]);
__builtin_mma_xvf64ger (&acc7, rowB1, rowA[3]);
for (l = 1; l < temp; l++)
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we have a macro like this for https://github.com/OpenMathLib/OpenBLAS/blob/develop/kernel/power/sgemm_kernel_power10.c#L128 to reduce the number of lines?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. I will add the macro same as sgemm.

{
rowA = (vec_t *) & AO[l << 3];
rowB = *((__vector_pair *)((void *)&BO[l << 3]));
rowB1 = *((__vector_pair *)((void *)&BO[(l << 3) + 4]));
__builtin_mma_xvf64gerpp (&acc0, rowB, rowA[0]);
__builtin_mma_xvf64gerpp (&acc1, rowB1, rowA[0]);
__builtin_mma_xvf64gerpp (&acc2, rowB, rowA[1]);
__builtin_mma_xvf64gerpp (&acc3, rowB1, rowA[1]);
__builtin_mma_xvf64gerpp (&acc4, rowB, rowA[2]);
__builtin_mma_xvf64gerpp (&acc5, rowB1, rowA[2]);
__builtin_mma_xvf64gerpp (&acc6, rowB, rowA[3]);
__builtin_mma_xvf64gerpp (&acc7, rowB1, rowA[3]);
}
for (l = 1; l + 15 < temp; l += 16)
{
KERNEL (l);
KERNEL (l+1);
KERNEL (l+2);
KERNEL (l+3);
KERNEL (l+4);
KERNEL (l+5);
KERNEL (l+6);
KERNEL (l+7);
KERNEL (l+8);
KERNEL (l+9);
KERNEL (l+10);
KERNEL (l+11);
KERNEL (l+12);
KERNEL (l+13);
KERNEL (l+14);
KERNEL (l+15);
}
if ((temp - l) & 8)
{
KERNEL(l);
KERNEL(l+1);
KERNEL(l+2);
KERNEL(l+3);
KERNEL(l+4);
KERNEL(l+5);
KERNEL(l+6);
KERNEL(l+7);
l += 8;
}
if ((temp - l) & 4)
{
KERNEL(l);
KERNEL(l+1);
KERNEL(l+2);
KERNEL(l+3);
l += 4;
}
if ((temp - l) & 2)
{
KERNEL(l);
KERNEL(l+1);
l += 2;
}
if ((temp - l) & 1)
{
KERNEL(l);
}
SAVE_ACC (&acc0, 0);
SAVE_ACC1 (&acc1, 0);
SAVE_ACC (&acc2, 2);
Expand Down
45 changes: 45 additions & 0 deletions kernel/power/dgemv_n_microk_power10.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,8 +25,53 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define HAVE_KERNEL_4x2 1
#define HAVE_KERNEL_4x1 1
#define HAVE_KERNEL_4x4 1

static void dgemv_kernel_4x2(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *xo, FLOAT *y, FLOAT alpha)
{
FLOAT x0,x1;
x0 = xo[0] * alpha;
x1 = xo[1] * alpha;
__vector double v_x0 = {x0,x0};
__vector double v_x1 = {x1,x1};
__vector double* v_y =(__vector double*)y;
__vector double* va0 = (__vector double*)a0;
__vector double* va1 = (__vector double*)a1;
for (int i=0; i< n/2; i+=2)
{

v_y[i]+= va0[i] * v_x0 + va1[i] * v_x1;
v_y[i+1]+= va0[i+1] * v_x0 + va1[i+1] * v_x1;

}


}

static void dgemv_kernel_4x1(BLASLONG n, FLOAT *a0, FLOAT *xo, FLOAT *y, FLOAT alpha)
{
BLASLONG i;
FLOAT x[1] __attribute__ ((aligned (16)));

FLOAT x0,x1;
x0 = xo[0] * alpha;

__vector double v_x0 = {x0,x0};
__vector double* v_y =(__vector double*)y;
__vector double* va0 = (__vector double*)a0;
for (int i=0; i< n/2; i+=2)
{

v_y[i]+= va0[i] * v_x0 ;
v_y[i+1]+= va0[i+1] * v_x0 ;

}

}


static void dgemv_kernel_4x4 (long n, double *ap, long lda, double *x, double *y, double alpha)
{
double *a0;
Expand Down
Loading