1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380
|
/*
BLIS
An object-based framework for developing high-performance BLAS-like
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2018 - 2019, 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(s) of the copyright holder(s) 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"
#define FUNCPTR_T gemm_fp
typedef void (*FUNCPTR_T)
(
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
);
static FUNCPTR_T GENARRAY(ftypes,gemm_ker_var2rr);
//
// -- Macrokernel functions for round-robin partitioning -----------------------
//
void bli_gemm_ker_var2rr
(
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;
FUNCPTR_T 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 );
// If 1m is being employed on a column- or row-stored matrix with a
// real-valued beta, we can use the real domain macro-kernel, which
// eliminates a little overhead associated with the 1m virtual
// micro-kernel.
if ( bli_is_1m_packed( schema_a ) )
{
bli_l3_ind_recast_1m_params
(
dt_exec,
schema_a,
c,
m, n, k,
pd_a, ps_a,
pd_b, ps_b,
rs_c, cs_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 PASTEMAC(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 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_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 ) \
{ \
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_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_rr( 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_var2rr: b1", k, NR, b1, NR, 1, "%4.1f", "" ); \
PASTEMAC(ch,fprintm)( stdout, "gemm_ker_var2rr: a1", MR, k, a1, 1, MR, "%4.1f", "" ); \
PASTEMAC(ch,fprintm)( stdout, "gemm_ker_var2rr: c after", m_cur, n_cur, c11, rs_c, cs_c, "%4.1f", "" ); \
*/ \
}
INSERT_GENTFUNC_BASIC0( gemm_ker_var2rr )
|