amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-20 07:38:53 +00:00
Code Issues Packages Projects Releases Wiki Activity

Added ability to handle non unit incx in GEMV transpose kernel. (#145)

Browse Source 
- GEMV transpose kernels lack ability to compute directly on non-unit stride inputs.
- This limitation is stopping libflame to use blis kernel directly instead of going through framework.
- Added ability to handle non-unit incx in the kernel by packing x into a temporary buffer.

AMD-Internal: [CPUPL-6903]
This commit is contained in:
Sharma, Shubham
2025-08-20 23:33:53 +05:30
committed by GitHub
parent 67616752c3
commit 805f36965d
2 changed files with 79 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
gtestsuite/testsuite/ukr/gemv/dgemv/dgemv_ukr.cpp
View File

@@ -456,8 +456,8 @@ INSTANTIATE_TEST_SUITE_P(
::testing::Range( gtint_t(1), gtint_t(16), gtint_t(1)), // n
::testing::Values( double(0.0), double(1.0), double(2.0)), // alpha
::testing::Values( double(0.0), double(1.0), double(2.0)), // beta
::testing::Values(gtint_t(1)), // stride size for x (non-unit incx is handled by frame)
::testing::Values(gtint_t(1), gtint_t(3)), // stride size for y
::testing::Values(gtint_t(1), gtint_t(8)), // stride size for x
::testing::Values(gtint_t(1), gtint_t(3)), // stride size for y
::testing::Values(gtint_t(0), gtint_t(7)), // increment to the leading dim of a
::testing::Values(false, true) // is_memory_test
),

79
kernels/zen4/2/bli_gemv_t_zen4_int.c
View File

@@ -110,7 +110,7 @@ static dgemv_ker_ft_conja n_ker_fp[8] =
dim_t n,
double* alpha,
double* a, inc_t inca, inc_t lda,
double* x, inc_t incx,
double* x0, inc_t incx0,
double* beta,
double* y, inc_t incy,
cntx_t* cntx
@@ -118,8 +118,75 @@ static dgemv_ker_ft_conja n_ker_fp[8] =
{
double* restrict a_buf = a;
double* restrict y_buf = y;
double* restrict x_buf = x;
double* x = x0; // optionally packed buffer
inc_t incx = incx0;
bool is_x_temp_buf_created = false;
mem_t mem_bufX;
rntm_t rntm_l;
if (incx != 1)
{
/*
Initialize mem pool buffer to NULL and size to 0
"buf" and "size" fields are assigned once memory
is allocated from the pool in bli_pba_acquire_m().
This will ensure bli_mem_is_alloc() will be passed on
an allocated memory if created or a NULL .
*/
mem_bufX.pblk.buf = NULL;
mem_bufX.pblk.block_size = 0;
mem_bufX.buf_type = 0;
mem_bufX.size = 0;
mem_bufX.pool = NULL;
// In order to get the buffer from pool via rntm_l access to memory broker
// is needed.Following are initializations for rntm_l.
bli_rntm_init_from_global(&rntm_l);
bli_rntm_set_num_threads_only(1, &rntm_l);
bli_pba_rntm_set_pba(&rntm_l);
//calculate the size required for m0 double elements in vector X.
size_t buffer_size = m * sizeof(double);
#ifdef BLIS_ENABLE_MEM_TRACING
printf("bli_dgemv_t_zen4_int(): get mem pool block for vector x\n");
#endif
// acquire a Buffer(m0*size(double)) from the memory broker
// and save the associated mem_t entry to mem_bufX.
bli_pba_acquire_m(&rntm_l,
buffer_size,
BLIS_BUFFER_FOR_B_PANEL,
&mem_bufX);
// Continue packing X if buffer memory is allocated.
if ( bli_mem_is_alloc(&mem_bufX) )
{
x = bli_mem_buffer(&mem_bufX);
// stride of vector x_buf =1
incx = 1;
// Invoke the COPYV function using the function pointer.
bli_dcopyv_zen4_asm
(
BLIS_NO_CONJUGATE,
m,
x0, incx0,
x, incx,
cntx
);
// Set x is packed as the memory allocation was successful
// and contents have been copied to a temp buffer.
is_x_temp_buf_created = TRUE;
}
}
double* x_buf = x; // temporary buffer
// i denotes the number of rows completed
// The variable 'j' is used to denote the number of rows that have been completed.
dim_t i = 0, j = 0;
@@ -1034,6 +1101,14 @@ static dgemv_ker_ft_conja n_ker_fp[8] =
cntx
);
}
if (is_x_temp_buf_created)
{
#ifdef BLIS_ENABLE_MEM_TRACING
printf("bli_dgemv_t_zen4_int(): releasing mem pool block for vector x\n");
#endif
// Return the buffer to pool
bli_pba_release(&rntm_l, &mem_bufX);
}
return;
}