amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 17:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix for DDOT Multi-thread implementation

Browse Source 
- In the existing implementation, when the actual number of threads
  spawned is different from the indicated number of threads for the
  parallel region, partial job is being executed.
- Fix added to identify actual number of threads spawned and
  allocate the work load to single thread in case of discrepancy
  in the number of threads spawned vs indicated.

AMD-Internal: [SWLCSG-1659]
Change-Id: I882d0af608009e502189a4469eb3483d659c2b3f
This commit is contained in:
Arnav Sharma
2022-11-18 16:27:13 +05:30
committed by Arnav Sharma
parent 11c42ce1d3
commit 69be3b0557
1 changed files with 65 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
frame/compat/bla_dot_amd.c
View File

@@ -300,6 +300,7 @@ double ddot_blis_impl
rntm_t rntm;
double* rho_temp = NULL;
dim_t nt, n0_per_thread, n0_rem, nt_pred;
dim_t i;
rho = 0;
// Initialize a local runtime with global settings.
bli_rntm_init_from_global(&rntm);
@@ -324,6 +325,13 @@ double ddot_blis_impl
return BLIS_NULL_POINTER;
rho_temp = bli_mem_buffer(&local_mem_buf);
if ( NULL == rho_temp ) return BLIS_NULL_POINTER;
// Initializing rho_temp array to 0
for ( i = 0; i < nt; i++ )
{
rho_temp[i] = 0;
}
#ifdef AOCL_DYNAMIC
// Calculate the optimal number of threads required
// based on input dimension. These conditions are taken
@@ -340,43 +348,72 @@ double ddot_blis_impl
n0_per_thread = n0 / nt;
n0_rem = n0 % nt;
// Multithreading Implementation
#pragma omp parallel num_threads(nt)
{
// Getting the actual number of threads that are spawned.
dim_t nt_real = omp_get_num_threads();
dim_t t_id = omp_get_thread_num();
// The following conditions handle the optimal distribution of
// load among the threads.
// Say we have n0 = 50 & nt = 4.
// So we get 12 ( n0 / nt ) elements per thread along with 2
// remaining elements. Each of these remaining elements is given
// to the last threads, respectively.
// So, t0, t1, t2 and t3 gets 12, 12, 13 and 13 elements,
// respectively.
dim_t npt, offset;
if ( t_id < ( nt - n0_rem ) )
// The actual number of threads spawned might be different
// from the predicted number of threads for which this parallel
// region is being generated. Thus, in such a case we are
// falling back to the Single-Threaded call.
if ( nt_real != nt )
{
npt = n0_per_thread;
offset = t_id * npt;
// More than one thread can still be spawned but since we
// are falling back to the ST call, we are
// calling the kernel from thread 0 only.
if ( t_id == 0 )
{
bli_ddotv_zen_int10
(
BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
n0,
x0, incx0,
y0, incy0,
rho_temp,
NULL
);
}
}
else
{
npt = n0_per_thread + 1;
offset = ( ( t_id * n0_per_thread ) +
( t_id - ( nt - n0_rem ) ) );
// The following conditions handle the optimal distribution
// of load among the threads.
// Say we have n0 = 50 & nt = 4.
// So we get 12 ( n0 / nt ) elements per thread along with 2
// remaining elements. Each of these remaining elements is
// given to the last threads, respectively.
// So, t0, t1, t2 and t3 gets 12, 12, 13 and 13 elements,
// respectively.
dim_t npt, offset;
if ( t_id < ( nt - n0_rem ) )
{
npt = n0_per_thread;
offset = t_id * npt;
}
else
{
npt = n0_per_thread + 1;
offset = ( ( t_id * n0_per_thread ) +
( t_id - ( nt - n0_rem ) ) );
}
bli_ddotv_zen_int10
(
BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
npt,
x0 + ( offset * incx0 ), incx0,
y0 + ( offset * incy0 ), incy0,
rho_temp + t_id,
NULL
);
}
bli_ddotv_zen_int10
(
BLIS_NO_CONJUGATE,
BLIS_NO_CONJUGATE,
npt,
x0 + ( offset * incx0 ), incx0,
y0 + ( offset * incy0 ), incy0,
rho_temp + t_id,
NULL
);
}
// Accumulating the nt threads output
for ( int i = 0; i < nt; i++ )
// Accumulating the nt thread outputs to rho
for ( i = 0; i < nt; i++ )
rho += rho_temp[i];
// Releasing the allocated memory