mirror of
https://github.com/amd/blis.git
synced 2026-04-26 18:48:51 +00:00
5e54f46ccb
Details: - Added a 'template' configuration, which contains stub implementations of the level 1, 1f, and 3 kernels with one datatype implemented in C for each, with lots of in-file comments and documentation. - Modified some variable/parameter names for some 1/1f operations. (e.g. renaming vector length parameter from m to n.) - Moved level-1f fusing factors from axpyf, dotxf, and dotxaxpyf header files to bli_kernel.h. - Modifed test suite to print out fusing factors for axpyf, dotxf, and dotxaxpyf, as well as the default fusing factor (which are all equal in the reference and template implementations). - Cleaned up some sloppiness in the level-1f unb_var1.c files whereby these reference variants were implemented in terms of front-end routines rather that directly in terms of the kernels. (For example, axpy2v was implemented as two calls to axpyv rather than two calls to AXPYV_KERNEL.) - Changed the interface to dotxf so that it matches that of axpyf, in that A is assumed to be m x b_n in both cases, and for dotxf A is actually used as A^T. - Minor variable naming and comment changes to reference micro-kernels in frame/3/gemm/ukernels and frame/3/trsm/ukernels.
291 lines
10 KiB
C
291 lines
10 KiB
C
/*
|
|
|
|
BLIS
|
|
An object-based framework for developing high-performance BLAS-like
|
|
libraries.
|
|
|
|
Copyright (C) 2013, The University of Texas
|
|
|
|
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 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"
|
|
|
|
|
|
|
|
void bli_sgemm_opt_mxn(
|
|
dim_t k,
|
|
float* restrict alpha,
|
|
float* restrict a,
|
|
float* restrict b,
|
|
float* restrict beta,
|
|
float* restrict c, inc_t rs_c, inc_t cs_c,
|
|
float* restrict a_next,
|
|
float* restrict b_next
|
|
)
|
|
{
|
|
/* Just call the reference implementation. */
|
|
bli_sgemm_ref_mxn( k,
|
|
alpha,
|
|
a,
|
|
b,
|
|
beta,
|
|
c, rs_c, cs_c,
|
|
a_next,
|
|
b_next );
|
|
}
|
|
|
|
|
|
|
|
void bli_dgemm_opt_mxn(
|
|
dim_t k,
|
|
double* restrict alpha,
|
|
double* restrict a,
|
|
double* restrict b,
|
|
double* restrict beta,
|
|
double* restrict c, inc_t rs_c, inc_t cs_c,
|
|
double* restrict a_next,
|
|
double* restrict b_next
|
|
)
|
|
{
|
|
/*
|
|
Template gemm micro-kernel implementation
|
|
|
|
This function contains a template implementation for a double-precision
|
|
real micro-kernel, coded in C, which can serve as the starting point for
|
|
one to write an optimized micro-kernel on an arbitrary architecture. (We
|
|
show a template implementation for only double-precision real because
|
|
the templates for the other three floating-point types would be nearly
|
|
identical.)
|
|
|
|
This micro-kernel performs a matrix-matrix multiplication of the form:
|
|
|
|
C := beta * C + alpha * A * B
|
|
|
|
where A is MR x k, B is k x NR, C is MR x NR, and alpha and beta are
|
|
scalars.
|
|
|
|
Parameters:
|
|
|
|
- k: The number of columns of A and rows of B.
|
|
- alpha: The address of a scalar to the A*B product.
|
|
- a: The address of a micro-panel of matrix A of dimension MR x k,
|
|
stored by columns.
|
|
- b: The address of a micro-panel of matrix B of dimension k x NR,
|
|
stored by rows.
|
|
- beta: The address of a scalar to the input value of matrix C.
|
|
- c: The address of a block of matrix C of dimension MR x NR,
|
|
stored according to rs_c and cs_c.
|
|
- rs_c: The row stride of matrix C (ie: the distance to the next row,
|
|
in units of matrix elements).
|
|
- cs_c: The column stride of matrix C (ie: the distance to the next
|
|
column, in units of matrix elements).
|
|
- a_next: The address of the micro-panel of A that will be used the next
|
|
time the gemm micro-kernel will be called.
|
|
- b_next: The address of the micro-panel of B that will be used the next
|
|
time the gemm micro-kernel will be called.
|
|
|
|
The diagram below shows the packed micro-panel operands and how elements
|
|
of each would be stored when MR == NR == 4. (The hex digits indicate the
|
|
order of the elements in memory.) Note that the storage of C is not shown
|
|
since it is determined by the row and column strides of C.
|
|
|
|
c: a: b:
|
|
_______ ______________________ _______
|
|
| | |0 4 8 C | |0 1 2 3|
|
|
MR | | |1 5 9 D . . . | |4 5 6 7|
|
|
| | += |2 6 A E | |8 9 A B|
|
|
|_______| |3_7_B_F_______________| |C D E F|
|
|
| . |
|
|
NR k | . |
|
|
| . |
|
|
| |
|
|
| |
|
|
|_______|
|
|
|
|
NR
|
|
Here are a few things to consider:
|
|
|
|
- bli_?mr and bli_?nr give the MR and NR register blocksizes for the
|
|
datatype corresponding to the '?' character.
|
|
- bli_?packmr and bli_?packnr are usually equal to bli_?mr and bli_?nr,
|
|
respectively. (They are only not equal if the register blocksize
|
|
extensions are non-zero. See bli_config.h for more details.)
|
|
- You may assume that the addresses a and b are aligned according to
|
|
the alignment value BLIS_CONTIG_STRIDE_ALIGN_SIZE, as defined in
|
|
bli_config.h.
|
|
- Here, we use a local array, ab, as temporary accumulator elements as
|
|
we compute the a*b product. In an optimized micro-kernel, ab is held
|
|
in registers rather than memory.
|
|
- In column-major storage (or column storage), the "leading dimension"
|
|
of a matrix is equivalent to its column stride, and the row stride is
|
|
unit. In row-major storage (row storage), the "leading dimension" is
|
|
equivalent to the row stride and the column stride is unit.
|
|
- While all three loops are exposed in this template micro-kernel, the
|
|
loops over MR and NR typically disappear in an optimized code because
|
|
they are fully unrolled, leaving only the loop over k.
|
|
- Some optimized micro-kernels will need the loop over k to be unrolled
|
|
a few times (4x seems to be a common unrolling factor).
|
|
- a_next and b_next can be used to perform prefetching, if prefetching
|
|
is supported by the architecture. They may be safely ignored by the
|
|
micro-kernel implementation, though.
|
|
- If beta == 0.0 (or 0.0 + 0.0i for complex), then the micro-kernel
|
|
should NOT use it explicitly, as C may contain uninitialized memory
|
|
(including NaNs). This case should be detected and handled separately,
|
|
preferably by simply overwriting C with the alpha*A*B product. An
|
|
example of how to perform this "beta is zero" handling is included in
|
|
this template implementation.
|
|
|
|
For more info, please refer to the BLIS website and/or contact the
|
|
blis-devel mailing list.
|
|
|
|
-FGVZ
|
|
*/
|
|
const dim_t mr = bli_dmr;
|
|
const dim_t nr = bli_dnr;
|
|
|
|
const inc_t cs_a = bli_dpackmr;
|
|
|
|
const inc_t rs_b = bli_dpacknr;
|
|
|
|
const inc_t rs_ab = 1;
|
|
const inc_t cs_ab = bli_dmr;
|
|
|
|
dim_t l, j, i;
|
|
|
|
double ab[ bli_dmr *
|
|
bli_dnr ];
|
|
double* abij;
|
|
double ai, bj;
|
|
|
|
|
|
/* Initialize the accumulator elements in ab to zero. */
|
|
for ( i = 0; i < mr * nr; ++i )
|
|
{
|
|
bli_dset0s( *(ab + i) );
|
|
}
|
|
|
|
/* Perform a series of k rank-1 updates into ab. */
|
|
for ( l = 0; l < k; ++l )
|
|
{
|
|
abij = ab;
|
|
|
|
/* In an optimized implementation, these two loops over MR and NR
|
|
are typically fully unrolled. */
|
|
for ( j = 0; j < nr; ++j )
|
|
{
|
|
bj = *(b + j);
|
|
|
|
for ( i = 0; i < mr; ++i )
|
|
{
|
|
ai = *(a + i);
|
|
|
|
bli_ddots( ai, bj, *abij );
|
|
|
|
abij += rs_ab;
|
|
}
|
|
}
|
|
|
|
a += cs_a;
|
|
b += rs_b;
|
|
}
|
|
|
|
/* Scale each element of ab by alpha. */
|
|
for ( i = 0; i < mr * nr; ++i )
|
|
{
|
|
bli_dscals( *alpha, *(ab + i) );
|
|
}
|
|
|
|
/* If beta is zero, overwrite c with the scaled result in ab. Otherwise,
|
|
scale c by beta and then add the scaled result in ab. */
|
|
if ( bli_deq0( *beta ) )
|
|
{
|
|
/* c := ab */
|
|
bli_dcopys_mxn( mr,
|
|
nr,
|
|
ab, rs_ab, cs_ab,
|
|
c, rs_c, cs_c );
|
|
}
|
|
else
|
|
{
|
|
/* c := beta * c + ab */
|
|
bli_dxpbys_mxn( mr,
|
|
nr,
|
|
ab, rs_ab, cs_ab,
|
|
beta,
|
|
c, rs_c, cs_c );
|
|
}
|
|
}
|
|
|
|
|
|
|
|
void bli_cgemm_opt_mxn(
|
|
dim_t k,
|
|
scomplex* restrict alpha,
|
|
scomplex* restrict a,
|
|
scomplex* restrict b,
|
|
scomplex* restrict beta,
|
|
scomplex* restrict c, inc_t rs_c, inc_t cs_c,
|
|
scomplex* restrict a_next,
|
|
scomplex* restrict b_next
|
|
)
|
|
{
|
|
/* Just call the reference implementation. */
|
|
bli_cgemm_ref_mxn( k,
|
|
alpha,
|
|
a,
|
|
b,
|
|
beta,
|
|
c, rs_c, cs_c,
|
|
a_next,
|
|
b_next );
|
|
}
|
|
|
|
|
|
|
|
void bli_zgemm_opt_mxn(
|
|
dim_t k,
|
|
dcomplex* restrict alpha,
|
|
dcomplex* restrict a,
|
|
dcomplex* restrict b,
|
|
dcomplex* restrict beta,
|
|
dcomplex* restrict c, inc_t rs_c, inc_t cs_c,
|
|
dcomplex* restrict a_next,
|
|
dcomplex* restrict b_next
|
|
)
|
|
{
|
|
/* Just call the reference implementation. */
|
|
bli_zgemm_ref_mxn( k,
|
|
alpha,
|
|
a,
|
|
b,
|
|
beta,
|
|
c, rs_c, cs_c,
|
|
a_next,
|
|
b_next );
|
|
}
|
|
|