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
|
/*
BLIS
An object-based framework for developing high-performance BLAS-like
libraries.
Copyright (C) 2014, The University of Texas at Austin
Copyright (C) 2020, Linaro Limited
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"
#include <arm_neon.h>
#if defined(__clang__)
#define PRAGMA_NOUNROLL _Pragma("nounroll")
#define PRAGMA_UNROLL_2 _Pragma("unroll 2")
#elif defined(__GNUC__)
#define PRAGMA_NOUNROLL _Pragma("GCC unroll 1")
#define PRAGMA_UNROLL_2 _Pragma("GCC unroll 2")
#else
#define PRAGMA_NOUNROLL
#define PRAGMA_UNROLL_2
#endif
void bli_dpackm_armv8a_int_6xk
(
conj_t conja,
pack_t schema,
dim_t cdim0,
dim_t k0,
dim_t k0_max,
double* restrict kappa,
double* restrict a, inc_t inca0, inc_t lda0,
double* restrict p, inc_t ldp0,
cntx_t* restrict cntx
)
{
// This is the panel dimension assumed by the packm kernel.
const dim_t mnr = 6;
// Typecast local copies of integers in case dim_t and inc_t are a
// different size than is expected by load instructions.
uint64_t k_iter = k0 / 2;
uint64_t k_left = k0 % 2;
double* a_loc = a;
double* p_loc = p;
// NOTE: For the purposes of the comments in this packm kernel, we
// interpret inca and lda as rs_a and cs_a, respectively, and similarly
// interpret ldp as cs_p (with rs_p implicitly unit). Thus, when reading
// this packm kernel, you should think of the operation as packing an
// m x n micropanel, where m and n are tiny and large, respectively, and
// where elements of each column of the packed matrix P are contiguous.
// (This packm kernel can still be used to pack micropanels of matrix B
// in a gemm operation.)
const uint64_t inca = inca0;
const uint64_t lda = lda0;
const uint64_t ldp = ldp0;
const bool gs = ( inca0 != 1 && lda0 != 1 );
// NOTE: If/when this kernel ever supports scaling by kappa within the
// assembly region, this constraint should be lifted.
const bool unitk = bli_deq1( *kappa );
// -------------------------------------------------------------------------
if ( cdim0 == mnr && !gs )
{
if ( unitk )
{
if ( inca == 1 )
{
// No need to use k-loops here.
// Simply let compiler to expand loops.
PRAGMA_UNROLL_2
for ( dim_t ik = k_iter * 2 + k_left; ik > 0; --ik )
{
float64x2_t v0 = vld1q_f64( a_loc + 0 );
float64x2_t v1 = vld1q_f64( a_loc + 2 );
float64x2_t v2 = vld1q_f64( a_loc + 4 );
vst1q_f64( p_loc + 0, v0 );
vst1q_f64( p_loc + 2, v1 );
vst1q_f64( p_loc + 4, v2 );
a_loc += lda;
p_loc += ldp;
}
}
else // if ( lda == 1 )
{
float64x2_t v0 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v1 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v2 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v3 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v4 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v5 = (float64x2_t)vdupq_n_u64( 0 );
PRAGMA_NOUNROLL
for ( ; k_iter > 0; --k_iter )
{
v0 = vld1q_f64( a_loc + inca * 0 );
v1 = vld1q_f64( a_loc + inca * 1 );
v2 = vld1q_f64( a_loc + inca * 2 );
v3 = vld1q_f64( a_loc + inca * 3 );
v4 = vld1q_f64( a_loc + inca * 4 );
v5 = vld1q_f64( a_loc + inca * 5 );
// In-register transpose.
float64x2_t vd0_1 = vtrn1q_f64( v0, v1 );
float64x2_t vd1_1 = vtrn1q_f64( v2, v3 );
float64x2_t vd2_1 = vtrn1q_f64( v4, v5 );
float64x2_t vd0_2 = vtrn2q_f64( v0, v1 );
float64x2_t vd1_2 = vtrn2q_f64( v2, v3 );
float64x2_t vd2_2 = vtrn2q_f64( v4, v5 );
vst1q_f64( p_loc + 0, vd0_1 );
vst1q_f64( p_loc + 2, vd1_1 );
vst1q_f64( p_loc + 4, vd2_1 );
p_loc += ldp;
vst1q_f64( p_loc + 0, vd0_2 );
vst1q_f64( p_loc + 2, vd1_2 );
vst1q_f64( p_loc + 4, vd2_2 );
p_loc += ldp;
a_loc += 2 * lda; // 2;
}
for ( ; k_left > 0; --k_left )
{
v0 = vld1q_lane_f64( a_loc + inca * 0, v0, 0 );
v0 = vld1q_lane_f64( a_loc + inca * 1, v0, 1 );
v1 = vld1q_lane_f64( a_loc + inca * 2, v1, 0 );
v1 = vld1q_lane_f64( a_loc + inca * 3, v1, 1 );
v2 = vld1q_lane_f64( a_loc + inca * 4, v2, 0 );
v2 = vld1q_lane_f64( a_loc + inca * 5, v2, 1 );
vst1q_f64( p_loc + 0, v0 );
vst1q_f64( p_loc + 2, v1 );
vst1q_f64( p_loc + 4, v2 );
p_loc += ldp;
a_loc += lda; // 1;
}
}
}
else // if ( !unitk )
{
float64x2_t vkappa = vld1q_dup_f64( kappa );
if ( inca == 1 )
{
// No need to use k-loops here.
// Simply let compiler to expand loops.
PRAGMA_UNROLL_2
for ( dim_t ik = k_iter * 2 + k_left; ik > 0; --ik )
{
float64x2_t v0 = vld1q_f64( a_loc + 0 );
float64x2_t v1 = vld1q_f64( a_loc + 2 );
float64x2_t v2 = vld1q_f64( a_loc + 4 );
// Scale by kappa.
v0 = vmulq_f64( v0, vkappa );
v1 = vmulq_f64( v1, vkappa );
v2 = vmulq_f64( v2, vkappa );
vst1q_f64( p_loc + 0, v0 );
vst1q_f64( p_loc + 2, v1 );
vst1q_f64( p_loc + 4, v2 );
a_loc += lda;
p_loc += ldp;
}
}
else // if ( lda == 1 )
{
float64x2_t v0 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v1 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v2 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v3 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v4 = (float64x2_t)vdupq_n_u64( 0 );
float64x2_t v5 = (float64x2_t)vdupq_n_u64( 0 );
PRAGMA_NOUNROLL
for ( ; k_iter > 0; --k_iter )
{
v0 = vld1q_f64( a_loc + inca * 0 );
v1 = vld1q_f64( a_loc + inca * 1 );
v2 = vld1q_f64( a_loc + inca * 2 );
v3 = vld1q_f64( a_loc + inca * 3 );
v4 = vld1q_f64( a_loc + inca * 4 );
v5 = vld1q_f64( a_loc + inca * 5 );
// Scale by kappa.
v0 = vmulq_f64( v0, vkappa );
v1 = vmulq_f64( v1, vkappa );
v2 = vmulq_f64( v2, vkappa );
v3 = vmulq_f64( v3, vkappa );
v4 = vmulq_f64( v4, vkappa );
v5 = vmulq_f64( v5, vkappa );
// In-register transpose.
float64x2_t vd0_1 = vtrn1q_f64( v0, v1 );
float64x2_t vd1_1 = vtrn1q_f64( v2, v3 );
float64x2_t vd2_1 = vtrn1q_f64( v4, v5 );
float64x2_t vd0_2 = vtrn2q_f64( v0, v1 );
float64x2_t vd1_2 = vtrn2q_f64( v2, v3 );
float64x2_t vd2_2 = vtrn2q_f64( v4, v5 );
vst1q_f64( p_loc + 0, vd0_1 );
vst1q_f64( p_loc + 2, vd1_1 );
vst1q_f64( p_loc + 4, vd2_1 );
p_loc += ldp;
vst1q_f64( p_loc + 0, vd0_2 );
vst1q_f64( p_loc + 2, vd1_2 );
vst1q_f64( p_loc + 4, vd2_2 );
p_loc += ldp;
a_loc += 2 * lda; // 2;
}
for ( ; k_left > 0; --k_left )
{
v0 = vld1q_lane_f64( a_loc + inca * 0, v0, 0 );
v0 = vld1q_lane_f64( a_loc + inca * 1, v0, 1 );
v1 = vld1q_lane_f64( a_loc + inca * 2, v1, 0 );
v1 = vld1q_lane_f64( a_loc + inca * 3, v1, 1 );
v2 = vld1q_lane_f64( a_loc + inca * 4, v2, 0 );
v2 = vld1q_lane_f64( a_loc + inca * 5, v2, 1 );
// Scale by kappa.
v0 = vmulq_f64( v0, vkappa );
v1 = vmulq_f64( v1, vkappa );
v2 = vmulq_f64( v2, vkappa );
vst1q_f64( p_loc + 0, v0 );
vst1q_f64( p_loc + 2, v1 );
vst1q_f64( p_loc + 4, v2 );
p_loc += ldp;
a_loc += lda; // 1;
}
}
}
}
else // if ( cdim0 < mnr || gs )
{
PASTEMAC(dscal2m,BLIS_TAPI_EX_SUF)
(
0,
BLIS_NONUNIT_DIAG,
BLIS_DENSE,
( trans_t )conja,
cdim0,
k0,
kappa,
a, inca0, lda0,
p, 1, ldp0,
cntx,
NULL
);
if ( cdim0 < mnr )
{
// Handle zero-filling along the "long" edge of the micropanel.
const dim_t i = cdim0;
const dim_t m_edge = mnr - cdim0;
const dim_t n_edge = k0_max;
double* restrict p_edge = p + (i )*1;
bli_dset0s_mxn
(
m_edge,
n_edge,
p_edge, 1, ldp
);
}
}
//bli_dfprintm( stdout, "packm 6xk ker: a_packed", cdim0, k0_max, p, 1, ldp0, "%5.2f", "" );
if ( k0 < k0_max )
{
// Handle zero-filling along the "short" (far) edge of the micropanel.
const dim_t j = k0;
const dim_t m_edge = mnr;
const dim_t n_edge = k0_max - k0;
double* restrict p_edge = p + (j )*ldp;
bli_dset0s_mxn
(
m_edge,
n_edge,
p_edge, 1, ldp
);
}
}
|
