amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-24 17:48:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Added option of slab or rr partitioning in jr/ir.

Browse Source 
Details:
- Updated existing macrokernel function names and definitions to
  explicitly use slab assignment of micropanels to threads, then created
  duplicate versions of macrokernels that explicitly use round-robin
  assignment instead of slab. NOTE: As in ac18949, trsm_r macrokernels
  were not substantially updated in this commit because they are
  currently disabled in bli_trsm_front.c.
- Updated existing packing function (in blk_packm_blk_var1.c) to
  explicitly use slab partitioning, and then duplicated for round-robin.
- Updated control tree initialization to use the appropriate macrokernel
  and packm function pointers depending on which method (slab or rr) was
  enabled at configure-time.
- Updated configure script to accept new --thread-part-jrir=[slab|rr]
  option (-m [slab|rr] for short), which allows the user to explicitly
  request either slab or round-robin assignment (partitioning) of
  micropanels to threads.
- Updated sandbox/ref99 according to above changes.
- Minor updates to build/add-copyright.py.
This commit is contained in:
Field G. Van Zee
2018-10-07 20:30:32 -05:00
parent 98e01ea04b
commit c92762ecdc
48 changed files with 7477 additions and 1394 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
sandbox/ref99/cntl/blx_gemm_cntl.c
View File

@@ -5,6 +5,7 @@
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2018, Advanced Micro Devices, Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
@@ -54,7 +55,28 @@ cntl_t* blx_gemmbp_cntl_create
pack_t schema_b
)
{
void* macro_kernel_p = blx_gemm_ker_var2;
void* macro_kernel_fp;
void* packa_fp;
void* packb_fp;
#ifdef BLIS_ENABLE_JRIR_SLAB
// Use the function pointers to the macrokernels that use slab
// assignment of micropanels to threads in the jr and ir loops.
macro_kernel_fp = blx_gemm_ker_var2sl;
packa_fp = bli_packm_blk_var1sl;
packb_fp = bli_packm_blk_var1sl;
#else // BLIS_ENABLE_JRIR_RR
// Use the function pointers to the macrokernels that use round-robin
// assignment of micropanels to threads in the jr and ir loops.
macro_kernel_fp = bli_gemm_ker_var2rr;
packa_fp = bli_packm_blk_var1rr;
packb_fp = bli_packm_blk_var1rr;
#endif
// Create two nodes for the macro-kernel.
cntl_t* gemm_cntl_bu_ke = blx_gemm_cntl_create_node
@@ -69,7 +91,7 @@ cntl_t* blx_gemmbp_cntl_create
(
family,
BLIS_NR, // not used by macro-kernel, but needed for bli_thrinfo_rgrow()
macro_kernel_p,
macro_kernel_fp,
gemm_cntl_bu_ke
);
@@ -77,7 +99,7 @@ cntl_t* blx_gemmbp_cntl_create
cntl_t* gemm_cntl_packa = blx_packm_cntl_create_node
(
blx_gemm_packa, // pack the left-hand operand
bli_packm_blk_var1,
packa_fp,
BLIS_MR,
BLIS_KR,
FALSE, // do NOT invert diagonal
@@ -101,7 +123,7 @@ cntl_t* blx_gemmbp_cntl_create
cntl_t* gemm_cntl_packb = blx_packm_cntl_create_node
(
blx_gemm_packb, // pack the right-hand operand
bli_packm_blk_var1,
packb_fp,
BLIS_KR,
BLIS_NR,
FALSE, // do NOT invert diagonal

28
sandbox/ref99/vars/blx_gemm_ker_var2.c → sandbox/ref99/vars/blx_gemm_ker_var2rr.c
View File

@@ -59,14 +59,14 @@ typedef void (*gemm_fp)
// Function pointer array for datatype-specific functions.
static gemm_fp ftypes[BLIS_NUM_FP_TYPES] =
{
PASTECH2(blx_,s,gemm_ker_var2),
PASTECH2(blx_,c,gemm_ker_var2),
PASTECH2(blx_,d,gemm_ker_var2),
PASTECH2(blx_,z,gemm_ker_var2)
PASTECH2(blx_,s,gemm_ker_var2rr),
PASTECH2(blx_,c,gemm_ker_var2rr),
PASTECH2(blx_,d,gemm_ker_var2rr),
PASTECH2(blx_,z,gemm_ker_var2rr)
};
void blx_gemm_ker_var2
void blx_gemm_ker_var2rr
(
obj_t* a,
obj_t* b,
@@ -272,8 +272,8 @@ void PASTECH2(blx_,ch,varname) \
dim_t jr_inc, ir_inc; \
\
/* Determine the thread range and increment for each thrinfo_t node. */ \
bli_thread_range_jrir( thread, n_iter, 1, FALSE, &jr_start, &jr_end, &jr_inc ); \
bli_thread_range_jrir( caucus, m_iter, 1, FALSE, &ir_start, &ir_end, &ir_inc ); \
bli_thread_range_jrir_rr( thread, n_iter, 1, FALSE, &jr_start, &jr_end, &jr_inc ); \
bli_thread_range_jrir_rr( caucus, m_iter, 1, FALSE, &ir_start, &ir_end, &ir_inc ); \
\
/* Loop over the n dimension (NR columns at a time). */ \
for ( j = jr_start; j < jr_end; j += jr_inc ) \
@@ -302,11 +302,11 @@ void PASTECH2(blx_,ch,varname) \
\
/* Compute the addresses of the next panels of A and B. */ \
a2 = bli_gemm_get_next_a_upanel( a1, rstep_a, ir_inc ); \
if ( bli_is_last_iter( i, ir_end, ir_tid, ir_nt ) ) \
if ( bli_is_last_iter_rr( i, ir_end, ir_tid, ir_nt ) ) \
{ \
a2 = a_cast; \
b2 = bli_gemm_get_next_b_upanel( b1, cstep_b, jr_inc ); \
if ( bli_is_last_iter( j, jr_end, jr_tid, jr_nt ) ) \
if ( bli_is_last_iter_rr( j, jr_end, jr_tid, jr_nt ) ) \
b2 = b_cast; \
} \
\
@@ -363,11 +363,11 @@ PASTEMAC(ch,fprintm)( stdout, "gemm_ker_var2: c after", m_cur, n_cur, c11, rs_c,
}
#if 0
GENTFUNC( float, s, gemm_ker_var2 )
GENTFUNC( double, d, gemm_ker_var2 )
GENTFUNC( scomplex, c, gemm_ker_var2 )
GENTFUNC( dcomplex, z, gemm_ker_var2 )
GENTFUNC( float, s, gemm_ker_var2rr )
GENTFUNC( double, d, gemm_ker_var2rr )
GENTFUNC( scomplex, c, gemm_ker_var2rr )
GENTFUNC( dcomplex, z, gemm_ker_var2rr )
#else
INSERT_GENTFUNC_BASIC0( gemm_ker_var2 )
INSERT_GENTFUNC_BASIC0( gemm_ker_var2rr )
#endif

373
sandbox/ref99/vars/blx_gemm_ker_var2sl.c Normal file
View File

@@ -0,0 +1,373 @@
/*
BLIS
An object-based framework for developing high-performance BLAS-like
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2018, Advanced Micro Devices, Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
- Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
- Neither the name of The University of Texas at Austin nor the names
of its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 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.
*/
#include "blis.h"
#include "blix.h"
// Function pointer type for datatype-specific functions.
typedef void (*gemm_fp)
(
pack_t schema_a,
pack_t schema_b,
dim_t m,
dim_t n,
dim_t k,
void* alpha,
void* a, inc_t cs_a, inc_t is_a,
dim_t pd_a, inc_t ps_a,
void* b, inc_t rs_b, inc_t is_b,
dim_t pd_b, inc_t ps_b,
void* beta,
void* c, inc_t rs_c, inc_t cs_c,
cntx_t* cntx,
rntm_t* rntm,
thrinfo_t* thread
);
// Function pointer array for datatype-specific functions.
static gemm_fp ftypes[BLIS_NUM_FP_TYPES] =
{
PASTECH2(blx_,s,gemm_ker_var2sl),
PASTECH2(blx_,c,gemm_ker_var2sl),
PASTECH2(blx_,d,gemm_ker_var2sl),
PASTECH2(blx_,z,gemm_ker_var2sl)
};
void blx_gemm_ker_var2sl
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
rntm_t* rntm,
cntl_t* cntl,
thrinfo_t* thread
)
{
num_t dt_exec = bli_obj_exec_dt( c );
pack_t schema_a = bli_obj_pack_schema( a );
pack_t schema_b = bli_obj_pack_schema( b );
dim_t m = bli_obj_length( c );
dim_t n = bli_obj_width( c );
dim_t k = bli_obj_width( a );
void* buf_a = bli_obj_buffer_at_off( a );
inc_t cs_a = bli_obj_col_stride( a );
inc_t is_a = bli_obj_imag_stride( a );
dim_t pd_a = bli_obj_panel_dim( a );
inc_t ps_a = bli_obj_panel_stride( a );
void* buf_b = bli_obj_buffer_at_off( b );
inc_t rs_b = bli_obj_row_stride( b );
inc_t is_b = bli_obj_imag_stride( b );
dim_t pd_b = bli_obj_panel_dim( b );
inc_t ps_b = bli_obj_panel_stride( b );
void* buf_c = bli_obj_buffer_at_off( c );
inc_t rs_c = bli_obj_row_stride( c );
inc_t cs_c = bli_obj_col_stride( c );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
gemm_fp f;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( &scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( schema_a,
schema_b,
m,
n,
k,
buf_alpha,
buf_a, cs_a, is_a,
pd_a, ps_a,
buf_b, rs_b, is_b,
pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
cntx,
rntm,
thread );
}
#undef GENTFUNC
#define GENTFUNC( ctype, ch, varname ) \
\
void PASTECH2(blx_,ch,varname) \
( \
pack_t schema_a, \
pack_t schema_b, \
dim_t m, \
dim_t n, \
dim_t k, \
void* alpha, \
void* a, inc_t cs_a, inc_t is_a, \
dim_t pd_a, inc_t ps_a, \
void* b, inc_t rs_b, inc_t is_b, \
dim_t pd_b, inc_t ps_b, \
void* beta, \
void* c, inc_t rs_c, inc_t cs_c, \
cntx_t* cntx, \
rntm_t* rntm, \
thrinfo_t* thread \
) \
{ \
const num_t dt = PASTEMAC(ch,type); \
\
/* Alias some constants to simpler names. */ \
const dim_t MR = pd_a; \
const dim_t NR = pd_b; \
/*const dim_t PACKMR = cs_a;*/ \
/*const dim_t PACKNR = rs_b;*/ \
\
/* Query the context for the micro-kernel address and cast it to its
function pointer type. */ \
PASTECH(ch,gemm_ukr_ft) \
gemm_ukr = bli_cntx_get_l3_vir_ukr_dt( dt, BLIS_GEMM_UKR, cntx ); \
\
/* Temporary C buffer for edge cases. Note that the strides of this
temporary buffer are set so that they match the storage of the
original C matrix. For example, if C is column-stored, ct will be
column-stored as well. */ \
ctype ct[ BLIS_STACK_BUF_MAX_SIZE \
/ sizeof( ctype ) ] \
__attribute__((aligned(BLIS_STACK_BUF_ALIGN_SIZE))); \
const bool_t col_pref = bli_cntx_l3_vir_ukr_prefers_cols_dt( dt, BLIS_GEMM_UKR, cntx ); \
const inc_t rs_ct = ( col_pref ? 1 : NR ); \
const inc_t cs_ct = ( col_pref ? MR : 1 ); \
\
ctype* restrict zero = PASTEMAC(ch,0); \
ctype* restrict a_cast = a; \
ctype* restrict b_cast = b; \
ctype* restrict c_cast = c; \
ctype* restrict alpha_cast = alpha; \
ctype* restrict beta_cast = beta; \
ctype* restrict b1; \
ctype* restrict c1; \
\
dim_t m_iter, m_left; \
dim_t n_iter, n_left; \
dim_t i, j; \
dim_t m_cur; \
dim_t n_cur; \
inc_t rstep_a; \
inc_t cstep_b; \
inc_t rstep_c, cstep_c; \
auxinfo_t aux; \
\
/*
Assumptions/assertions:
rs_a == 1
cs_a == PACKMR
pd_a == MR
ps_a == stride to next micro-panel of A
rs_b == PACKNR
cs_b == 1
pd_b == NR
ps_b == stride to next micro-panel of B
rs_c == (no assumptions)
cs_c == (no assumptions)
*/ \
\
/* If any dimension is zero, return immediately. */ \
if ( bli_zero_dim3( m, n, k ) ) return; \
\
/* Clear the temporary C buffer in case it has any infs or NaNs. */ \
PASTEMAC(ch,set0s_mxn)( MR, NR, \
ct, rs_ct, cs_ct ); \
\
/* Compute number of primary and leftover components of the m and n
dimensions. */ \
n_iter = n / NR; \
n_left = n % NR; \
\
m_iter = m / MR; \
m_left = m % MR; \
\
if ( n_left ) ++n_iter; \
if ( m_left ) ++m_iter; \
\
/* Determine some increments used to step through A, B, and C. */ \
rstep_a = ps_a; \
\
cstep_b = ps_b; \
\
rstep_c = rs_c * MR; \
cstep_c = cs_c * NR; \
\
/* Save the pack schemas of A and B to the auxinfo_t object. */ \
bli_auxinfo_set_schema_a( schema_a, &aux ); \
bli_auxinfo_set_schema_b( schema_b, &aux ); \
\
/* Save the imaginary stride of A and B to the auxinfo_t object. */ \
bli_auxinfo_set_is_a( is_a, &aux ); \
bli_auxinfo_set_is_b( is_b, &aux ); \
\
/* The 'thread' argument points to the thrinfo_t node for the 2nd (jr)
loop around the microkernel. Here we query the thrinfo_t node for the
1st (ir) loop around the microkernel. */ \
thrinfo_t* caucus = bli_thrinfo_sub_node( thread ); \
\
/* Query the number of threads and thread ids for each loop. */ \
dim_t jr_nt = bli_thread_n_way( thread ); \
dim_t jr_tid = bli_thread_work_id( thread ); \
dim_t ir_nt = bli_thread_n_way( caucus ); \
dim_t ir_tid = bli_thread_work_id( caucus ); \
\
dim_t jr_start, jr_end; \
dim_t ir_start, ir_end; \
dim_t jr_inc, ir_inc; \
\
/* Determine the thread range and increment for each thrinfo_t node. */ \
bli_thread_range_jrir_sl( thread, n_iter, 1, FALSE, &jr_start, &jr_end, &jr_inc ); \
bli_thread_range_jrir_sl( caucus, m_iter, 1, FALSE, &ir_start, &ir_end, &ir_inc ); \
\
/* Loop over the n dimension (NR columns at a time). */ \
for ( j = jr_start; j < jr_end; j += jr_inc ) \
{ \
ctype* restrict a1; \
ctype* restrict c11; \
ctype* restrict b2; \
\
b1 = b_cast + j * cstep_b; \
c1 = c_cast + j * cstep_c; \
\
n_cur = ( bli_is_not_edge_f( j, n_iter, n_left ) ? NR : n_left ); \
\
/* Initialize our next panel of B to be the current panel of B. */ \
b2 = b1; \
\
/* Loop over the m dimension (MR rows at a time). */ \
for ( i = ir_start; i < ir_end; i += ir_inc ) \
{ \
ctype* restrict a2; \
\
a1 = a_cast + i * rstep_a; \
c11 = c1 + i * rstep_c; \
\
m_cur = ( bli_is_not_edge_f( i, m_iter, m_left ) ? MR : m_left ); \
\
/* Compute the addresses of the next panels of A and B. */ \
a2 = bli_gemm_get_next_a_upanel( a1, rstep_a, ir_inc ); \
if ( bli_is_last_iter_sl( i, ir_end, ir_tid, ir_nt ) ) \
{ \
a2 = a_cast; \
b2 = bli_gemm_get_next_b_upanel( b1, cstep_b, jr_inc ); \
if ( bli_is_last_iter_sl( j, jr_end, jr_tid, jr_nt ) ) \
b2 = b_cast; \
} \
\
/* Save addresses of next panels of A and B to the auxinfo_t
object. */ \
bli_auxinfo_set_next_a( a2, &aux ); \
bli_auxinfo_set_next_b( b2, &aux ); \
\
/* Handle interior and edge cases separately. */ \
if ( m_cur == MR && n_cur == NR ) \
{ \
/* Invoke the gemm micro-kernel. */ \
gemm_ukr \
( \
k, \
alpha_cast, \
a1, \
b1, \
beta_cast, \
c11, rs_c, cs_c, \
&aux, \
cntx \
); \
} \
else \
{ \
/* Invoke the gemm micro-kernel. */ \
gemm_ukr \
( \
k, \
alpha_cast, \
a1, \
b1, \
zero, \
ct, rs_ct, cs_ct, \
&aux, \
cntx \
); \
\
/* Scale the bottom edge of C and add the result from above. */ \
PASTEMAC(ch,xpbys_mxn)( m_cur, n_cur, \
ct, rs_ct, cs_ct, \
beta_cast, \
c11, rs_c, cs_c ); \
} \
} \
} \
\
/*
PASTEMAC(ch,fprintm)( stdout, "gemm_ker_var2: b1", k, NR, b1, NR, 1, "%4.1f", "" ); \
PASTEMAC(ch,fprintm)( stdout, "gemm_ker_var2: a1", MR, k, a1, 1, MR, "%4.1f", "" ); \
PASTEMAC(ch,fprintm)( stdout, "gemm_ker_var2: c after", m_cur, n_cur, c11, rs_c, cs_c, "%4.1f", "" ); \
*/ \
}
#if 0
GENTFUNC( float, s, gemm_ker_var2sl )
GENTFUNC( double, d, gemm_ker_var2sl )
GENTFUNC( scomplex, c, gemm_ker_var2sl )
GENTFUNC( dcomplex, z, gemm_ker_var2sl )
#else
INSERT_GENTFUNC_BASIC0( gemm_ker_var2sl )
#endif

7
sandbox/ref99/vars/blx_gemm_var.h
View File

@@ -5,6 +5,7 @@
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2018, Advanced Micro Devices, Inc.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
@@ -57,7 +58,8 @@ GENPROT( gemm_blk_var3 )
GENPROT( gemm_packa )
GENPROT( gemm_packb )
GENPROT( gemm_ker_var2 )
GENPROT( gemm_ker_var2sl )
GENPROT( gemm_ker_var2rr )
//
// Prototype BLAS-like interfaces with void pointer operands.
@@ -85,5 +87,6 @@ void PASTECH2(blx_,ch,varname) \
thrinfo_t* thread \
);
INSERT_GENTPROT_BASIC0( gemm_ker_var2 )
INSERT_GENTPROT_BASIC0( gemm_ker_var2sl )
INSERT_GENTPROT_BASIC0( gemm_ker_var2rr )