1 |
|
/* |
2 |
|
|
3 |
|
BLIS |
4 |
|
An object-based framework for developing high-performance BLAS-like |
5 |
|
libraries. |
6 |
|
|
7 |
|
Copyright (C) 2014, The University of Texas at Austin |
8 |
|
|
9 |
|
Redistribution and use in source and binary forms, with or without |
10 |
|
modification, are permitted provided that the following conditions are |
11 |
|
met: |
12 |
|
- Redistributions of source code must retain the above copyright |
13 |
|
notice, this list of conditions and the following disclaimer. |
14 |
|
- Redistributions in binary form must reproduce the above copyright |
15 |
|
notice, this list of conditions and the following disclaimer in the |
16 |
|
documentation and/or other materials provided with the distribution. |
17 |
|
- Neither the name of The University of Texas at Austin nor the names |
18 |
|
of its contributors may be used to endorse or promote products |
19 |
|
derived from this software without specific prior written permission. |
20 |
|
|
21 |
|
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
22 |
|
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
23 |
|
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
24 |
|
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT |
25 |
|
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, |
26 |
|
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT |
27 |
|
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, |
28 |
|
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY |
29 |
|
THEORY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
30 |
|
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
31 |
|
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
32 |
|
|
33 |
|
*/ |
34 |
|
|
35 |
|
//#include "blis.h" |
36 |
|
#include <hmlp_internal.hpp> |
37 |
|
|
38 |
|
|
39 |
|
#define SGEMM_INPUT_GS_BETA_NZ \ |
40 |
|
"vmovlps (%%rcx ), %%xmm0, %%xmm0 \n\t" \ |
41 |
|
"vmovhps (%%rcx,%%rsi,1), %%xmm0, %%xmm0 \n\t" \ |
42 |
|
"vmovlps (%%rcx,%%rsi,2), %%xmm1, %%xmm1 \n\t" \ |
43 |
|
"vmovhps (%%rcx,%%r13 ), %%xmm1, %%xmm1 \n\t" \ |
44 |
|
"vshufps $0x88, %%xmm1, %%xmm0, %%xmm0 \n\t" \ |
45 |
|
"vmovlps (%%rcx,%%rsi,4), %%xmm2, %%xmm2 \n\t" \ |
46 |
|
"vmovhps (%%rcx,%%r15 ), %%xmm2, %%xmm2 \n\t" \ |
47 |
|
"vmovlps (%%rcx,%%r13,2), %%xmm1, %%xmm1 \n\t" \ |
48 |
|
"vmovhps (%%rcx,%%r10 ), %%xmm1, %%xmm1 \n\t" \ |
49 |
|
"vshufps $0x88, %%xmm1, %%xmm2, %%xmm2 \n\t" \ |
50 |
|
"vperm2f128 $0x20, %%ymm2, %%ymm0, %%ymm0 \n\t" |
51 |
|
|
52 |
|
#define SGEMM_OUTPUT_GS_BETA_NZ \ |
53 |
|
"vextractf128 $1, %%ymm0, %%xmm2 \n\t" \ |
54 |
|
"vmovss %%xmm0, (%%rcx ) \n\t" \ |
55 |
|
"vpermilps $0x39, %%xmm0, %%xmm1 \n\t" \ |
56 |
|
"vmovss %%xmm1, (%%rcx,%%rsi,1) \n\t" \ |
57 |
|
"vpermilps $0x39, %%xmm1, %%xmm0 \n\t" \ |
58 |
|
"vmovss %%xmm0, (%%rcx,%%rsi,2) \n\t" \ |
59 |
|
"vpermilps $0x39, %%xmm0, %%xmm1 \n\t" \ |
60 |
|
"vmovss %%xmm1, (%%rcx,%%r13 ) \n\t" \ |
61 |
|
"vmovss %%xmm2, (%%rcx,%%rsi,4) \n\t" \ |
62 |
|
"vpermilps $0x39, %%xmm2, %%xmm1 \n\t" \ |
63 |
|
"vmovss %%xmm1, (%%rcx,%%r15 ) \n\t" \ |
64 |
|
"vpermilps $0x39, %%xmm1, %%xmm2 \n\t" \ |
65 |
|
"vmovss %%xmm2, (%%rcx,%%r13,2) \n\t" \ |
66 |
|
"vpermilps $0x39, %%xmm2, %%xmm1 \n\t" \ |
67 |
|
"vmovss %%xmm1, (%%rcx,%%r10 ) \n\t" |
68 |
|
|
69 |
|
void bli_sgemm_asm_16x6 |
70 |
|
( |
71 |
|
dim_t k, |
72 |
|
float* restrict alpha, |
73 |
|
float* restrict a, |
74 |
|
float* restrict b, |
75 |
|
float* restrict beta, |
76 |
|
float* restrict c, inc_t rs_c, inc_t cs_c, |
77 |
|
//auxinfo_t* restrict data, |
78 |
|
//cntx_t* restrict cntx |
79 |
|
aux_s<float, float, float, float> *aux |
80 |
|
) |
81 |
|
{ |
82 |
|
//void* a_next = bli_auxinfo_next_a( data ); |
83 |
|
//void* b_next = bli_auxinfo_next_b( data ); |
84 |
|
|
85 |
|
dim_t k_iter = k / 4; |
86 |
|
dim_t k_left = k % 4; |
87 |
|
|
88 |
|
__asm__ volatile |
89 |
|
( |
90 |
|
" \n\t" |
91 |
|
"vzeroall \n\t" // zero all xmm/ymm registers. |
92 |
|
" \n\t" |
93 |
|
" \n\t" |
94 |
|
"movq %2, %%rax \n\t" // load address of a. |
95 |
|
"movq %3, %%rbx \n\t" // load address of b. |
96 |
|
//"movq %9, %%r15 \n\t" // load address of b_next. |
97 |
|
" \n\t" |
98 |
|
"addq $32 * 4, %%rax \n\t" |
99 |
|
" \n\t" // initialize loop by pre-loading |
100 |
|
"vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
101 |
|
"vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
102 |
|
" \n\t" |
103 |
|
"movq %6, %%rcx \n\t" // load address of c |
104 |
|
"movq %8, %%rdi \n\t" // load cs_c |
105 |
|
"leaq (,%%rdi,4), %%rdi \n\t" // cs_c *= sizeof(float) |
106 |
|
" \n\t" |
107 |
|
"leaq (%%rdi,%%rdi,2), %%r13 \n\t" // r13 = 3*cs_c; |
108 |
|
"leaq (%%rcx,%%r13,1), %%rdx \n\t" // rdx = c + 3*cs_c; |
109 |
|
"prefetcht0 7 * 8(%%rcx) \n\t" // prefetch c + 0*cs_c |
110 |
|
"prefetcht0 7 * 8(%%rcx,%%rdi) \n\t" // prefetch c + 1*cs_c |
111 |
|
"prefetcht0 7 * 8(%%rcx,%%rdi,2) \n\t" // prefetch c + 2*cs_c |
112 |
|
"prefetcht0 7 * 8(%%rdx) \n\t" // prefetch c + 3*cs_c |
113 |
|
"prefetcht0 7 * 8(%%rdx,%%rdi) \n\t" // prefetch c + 4*cs_c |
114 |
|
"prefetcht0 7 * 8(%%rdx,%%rdi,2) \n\t" // prefetch c + 5*cs_c |
115 |
|
" \n\t" |
116 |
|
" \n\t" |
117 |
|
" \n\t" |
118 |
|
" \n\t" |
119 |
|
"movq %0, %%rsi \n\t" // i = k_iter; |
120 |
|
"testq %%rsi, %%rsi \n\t" // check i via logical AND. |
121 |
|
"je .SCONSIDKLEFT \n\t" // if i == 0, jump to code that |
122 |
|
" \n\t" // contains the k_left loop. |
123 |
|
" \n\t" |
124 |
|
" \n\t" |
125 |
|
".SLOOPKITER: \n\t" // MAIN LOOP |
126 |
|
" \n\t" |
127 |
|
" \n\t" |
128 |
|
" \n\t" // iteration 0 |
129 |
|
"prefetcht0 128 * 4(%%rax) \n\t" |
130 |
|
" \n\t" |
131 |
|
"vbroadcastss 0 * 4(%%rbx), %%ymm2 \n\t" |
132 |
|
"vbroadcastss 1 * 4(%%rbx), %%ymm3 \n\t" |
133 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
134 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
135 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
136 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
137 |
|
" \n\t" |
138 |
|
"vbroadcastss 2 * 4(%%rbx), %%ymm2 \n\t" |
139 |
|
"vbroadcastss 3 * 4(%%rbx), %%ymm3 \n\t" |
140 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
141 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
142 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
143 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
144 |
|
" \n\t" |
145 |
|
"vbroadcastss 4 * 4(%%rbx), %%ymm2 \n\t" |
146 |
|
"vbroadcastss 5 * 4(%%rbx), %%ymm3 \n\t" |
147 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
148 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
149 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
150 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
151 |
|
" \n\t" |
152 |
|
"vmovaps -2 * 32(%%rax), %%ymm0 \n\t" |
153 |
|
"vmovaps -1 * 32(%%rax), %%ymm1 \n\t" |
154 |
|
" \n\t" |
155 |
|
" \n\t" // iteration 1 |
156 |
|
"vbroadcastss 6 * 4(%%rbx), %%ymm2 \n\t" |
157 |
|
"vbroadcastss 7 * 4(%%rbx), %%ymm3 \n\t" |
158 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
159 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
160 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
161 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
162 |
|
" \n\t" |
163 |
|
"vbroadcastss 8 * 4(%%rbx), %%ymm2 \n\t" |
164 |
|
"vbroadcastss 9 * 4(%%rbx), %%ymm3 \n\t" |
165 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
166 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
167 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
168 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
169 |
|
" \n\t" |
170 |
|
"vbroadcastss 10 * 4(%%rbx), %%ymm2 \n\t" |
171 |
|
"vbroadcastss 11 * 4(%%rbx), %%ymm3 \n\t" |
172 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
173 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
174 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
175 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
176 |
|
" \n\t" |
177 |
|
"vmovaps 0 * 32(%%rax), %%ymm0 \n\t" |
178 |
|
"vmovaps 1 * 32(%%rax), %%ymm1 \n\t" |
179 |
|
" \n\t" |
180 |
|
" \n\t" // iteration 2 |
181 |
|
"prefetcht0 152 * 4(%%rax) \n\t" |
182 |
|
" \n\t" |
183 |
|
"vbroadcastss 12 * 4(%%rbx), %%ymm2 \n\t" |
184 |
|
"vbroadcastss 13 * 4(%%rbx), %%ymm3 \n\t" |
185 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
186 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
187 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
188 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
189 |
|
" \n\t" |
190 |
|
"vbroadcastss 14 * 4(%%rbx), %%ymm2 \n\t" |
191 |
|
"vbroadcastss 15 * 4(%%rbx), %%ymm3 \n\t" |
192 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
193 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
194 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
195 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
196 |
|
" \n\t" |
197 |
|
"vbroadcastss 16 * 4(%%rbx), %%ymm2 \n\t" |
198 |
|
"vbroadcastss 17 * 4(%%rbx), %%ymm3 \n\t" |
199 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
200 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
201 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
202 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
203 |
|
" \n\t" |
204 |
|
"vmovaps 2 * 32(%%rax), %%ymm0 \n\t" |
205 |
|
"vmovaps 3 * 32(%%rax), %%ymm1 \n\t" |
206 |
|
" \n\t" |
207 |
|
" \n\t" // iteration 3 |
208 |
|
"vbroadcastss 18 * 4(%%rbx), %%ymm2 \n\t" |
209 |
|
"vbroadcastss 19 * 4(%%rbx), %%ymm3 \n\t" |
210 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
211 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
212 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
213 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
214 |
|
" \n\t" |
215 |
|
"vbroadcastss 20 * 4(%%rbx), %%ymm2 \n\t" |
216 |
|
"vbroadcastss 21 * 4(%%rbx), %%ymm3 \n\t" |
217 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
218 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
219 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
220 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
221 |
|
" \n\t" |
222 |
|
"vbroadcastss 22 * 4(%%rbx), %%ymm2 \n\t" |
223 |
|
"vbroadcastss 23 * 4(%%rbx), %%ymm3 \n\t" |
224 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
225 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
226 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
227 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
228 |
|
" \n\t" |
229 |
|
"addq $4 * 16 * 4, %%rax \n\t" // a += 4*16 (unroll x mr) |
230 |
|
"addq $4 * 6 * 4, %%rbx \n\t" // b += 4*6 (unroll x nr) |
231 |
|
" \n\t" |
232 |
|
"vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
233 |
|
"vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
234 |
|
" \n\t" |
235 |
|
" \n\t" |
236 |
|
"decq %%rsi \n\t" // i -= 1; |
237 |
|
"jne .SLOOPKITER \n\t" // iterate again if i != 0. |
238 |
|
" \n\t" |
239 |
|
" \n\t" |
240 |
|
" \n\t" |
241 |
|
" \n\t" |
242 |
|
" \n\t" |
243 |
|
" \n\t" |
244 |
|
".SCONSIDKLEFT: \n\t" |
245 |
|
" \n\t" |
246 |
|
"movq %1, %%rsi \n\t" // i = k_left; |
247 |
|
"testq %%rsi, %%rsi \n\t" // check i via logical AND. |
248 |
|
"je .SPOSTACCUM \n\t" // if i == 0, we're done; jump to end. |
249 |
|
" \n\t" // else, we prepare to enter k_left loop. |
250 |
|
" \n\t" |
251 |
|
" \n\t" |
252 |
|
".SLOOPKLEFT: \n\t" // EDGE LOOP |
253 |
|
" \n\t" |
254 |
|
"prefetcht0 128 * 4(%%rax) \n\t" |
255 |
|
" \n\t" |
256 |
|
"vbroadcastss 0 * 4(%%rbx), %%ymm2 \n\t" |
257 |
|
"vbroadcastss 1 * 4(%%rbx), %%ymm3 \n\t" |
258 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
259 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
260 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
261 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
262 |
|
" \n\t" |
263 |
|
"vbroadcastss 2 * 4(%%rbx), %%ymm2 \n\t" |
264 |
|
"vbroadcastss 3 * 4(%%rbx), %%ymm3 \n\t" |
265 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
266 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
267 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
268 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
269 |
|
" \n\t" |
270 |
|
"vbroadcastss 4 * 4(%%rbx), %%ymm2 \n\t" |
271 |
|
"vbroadcastss 5 * 4(%%rbx), %%ymm3 \n\t" |
272 |
|
"vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
273 |
|
"vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
274 |
|
"vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
275 |
|
"vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
276 |
|
" \n\t" |
277 |
|
"addq $1 * 16 * 4, %%rax \n\t" // a += 1*16 (unroll x mr) |
278 |
|
"addq $1 * 6 * 4, %%rbx \n\t" // b += 1*6 (unroll x nr) |
279 |
|
" \n\t" |
280 |
|
"vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
281 |
|
"vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
282 |
|
" \n\t" |
283 |
|
" \n\t" |
284 |
|
"decq %%rsi \n\t" // i -= 1; |
285 |
|
"jne .SLOOPKLEFT \n\t" // iterate again if i != 0. |
286 |
|
" \n\t" |
287 |
|
" \n\t" |
288 |
|
" \n\t" |
289 |
|
".SPOSTACCUM: \n\t" |
290 |
|
" \n\t" |
291 |
|
" \n\t" |
292 |
|
" \n\t" |
293 |
|
" \n\t" |
294 |
|
"movq %4, %%rax \n\t" // load address of alpha |
295 |
|
"movq %5, %%rbx \n\t" // load address of beta |
296 |
|
"vbroadcastss (%%rax), %%ymm0 \n\t" // load alpha and duplicate |
297 |
|
"vbroadcastss (%%rbx), %%ymm3 \n\t" // load beta and duplicate |
298 |
|
" \n\t" |
299 |
|
"vmulps %%ymm0, %%ymm4, %%ymm4 \n\t" // scale by alpha |
300 |
|
"vmulps %%ymm0, %%ymm5, %%ymm5 \n\t" |
301 |
|
"vmulps %%ymm0, %%ymm6, %%ymm6 \n\t" |
302 |
|
"vmulps %%ymm0, %%ymm7, %%ymm7 \n\t" |
303 |
|
"vmulps %%ymm0, %%ymm8, %%ymm8 \n\t" |
304 |
|
"vmulps %%ymm0, %%ymm9, %%ymm9 \n\t" |
305 |
|
"vmulps %%ymm0, %%ymm10, %%ymm10 \n\t" |
306 |
|
"vmulps %%ymm0, %%ymm11, %%ymm11 \n\t" |
307 |
|
"vmulps %%ymm0, %%ymm12, %%ymm12 \n\t" |
308 |
|
"vmulps %%ymm0, %%ymm13, %%ymm13 \n\t" |
309 |
|
"vmulps %%ymm0, %%ymm14, %%ymm14 \n\t" |
310 |
|
"vmulps %%ymm0, %%ymm15, %%ymm15 \n\t" |
311 |
|
" \n\t" |
312 |
|
" \n\t" |
313 |
|
" \n\t" |
314 |
|
" \n\t" |
315 |
|
" \n\t" |
316 |
|
" \n\t" |
317 |
|
"movq %7, %%rsi \n\t" // load rs_c |
318 |
|
"leaq (,%%rsi,4), %%rsi \n\t" // rsi = rs_c * sizeof(float) |
319 |
|
" \n\t" |
320 |
|
"leaq (%%rcx,%%rsi,8), %%rdx \n\t" // load address of c + 8*rs_c; |
321 |
|
" \n\t" |
322 |
|
"leaq (%%rsi,%%rsi,2), %%r13 \n\t" // r13 = 3*rs_c; |
323 |
|
"leaq (%%rsi,%%rsi,4), %%r15 \n\t" // r15 = 5*rs_c; |
324 |
|
"leaq (%%r13,%%rsi,4), %%r10 \n\t" // r10 = 7*rs_c; |
325 |
|
" \n\t" |
326 |
|
" \n\t" |
327 |
|
" \n\t" // now avoid loading C if beta == 0 |
328 |
|
" \n\t" |
329 |
|
"vxorps %%ymm0, %%ymm0, %%ymm0 \n\t" // set ymm0 to zero. |
330 |
|
"vucomiss %%xmm0, %%xmm3 \n\t" // set ZF if beta == 0. |
331 |
|
"je .SBETAZERO \n\t" // if ZF = 1, jump to beta == 0 case |
332 |
|
" \n\t" |
333 |
|
" \n\t" |
334 |
|
"cmpq $4, %%rsi \n\t" // set ZF if (4*rs_c) == 4. |
335 |
|
"jz .SCOLSTORED \n\t" // jump to column storage case |
336 |
|
" \n\t" |
337 |
|
" \n\t" |
338 |
|
" \n\t" |
339 |
|
".SGENSTORED: \n\t" |
340 |
|
" \n\t" |
341 |
|
" \n\t" |
342 |
|
SGEMM_INPUT_GS_BETA_NZ |
343 |
|
"vfmadd213ps %%ymm4, %%ymm3, %%ymm0 \n\t" |
344 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
345 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
346 |
|
" \n\t" |
347 |
|
" \n\t" |
348 |
|
SGEMM_INPUT_GS_BETA_NZ |
349 |
|
"vfmadd213ps %%ymm6, %%ymm3, %%ymm0 \n\t" |
350 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
351 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
352 |
|
" \n\t" |
353 |
|
" \n\t" |
354 |
|
SGEMM_INPUT_GS_BETA_NZ |
355 |
|
"vfmadd213ps %%ymm8, %%ymm3, %%ymm0 \n\t" |
356 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
357 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
358 |
|
" \n\t" |
359 |
|
" \n\t" |
360 |
|
SGEMM_INPUT_GS_BETA_NZ |
361 |
|
"vfmadd213ps %%ymm10, %%ymm3, %%ymm0 \n\t" |
362 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
363 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
364 |
|
" \n\t" |
365 |
|
" \n\t" |
366 |
|
SGEMM_INPUT_GS_BETA_NZ |
367 |
|
"vfmadd213ps %%ymm12, %%ymm3, %%ymm0 \n\t" |
368 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
369 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
370 |
|
" \n\t" |
371 |
|
" \n\t" |
372 |
|
SGEMM_INPUT_GS_BETA_NZ |
373 |
|
"vfmadd213ps %%ymm14, %%ymm3, %%ymm0 \n\t" |
374 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
375 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
376 |
|
" \n\t" |
377 |
|
" \n\t" |
378 |
|
"movq %%rdx, %%rcx \n\t" // rcx = c + 8*rs_c |
379 |
|
" \n\t" |
380 |
|
" \n\t" |
381 |
|
SGEMM_INPUT_GS_BETA_NZ |
382 |
|
"vfmadd213ps %%ymm5, %%ymm3, %%ymm0 \n\t" |
383 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
384 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
385 |
|
" \n\t" |
386 |
|
" \n\t" |
387 |
|
SGEMM_INPUT_GS_BETA_NZ |
388 |
|
"vfmadd213ps %%ymm7, %%ymm3, %%ymm0 \n\t" |
389 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
390 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
391 |
|
" \n\t" |
392 |
|
" \n\t" |
393 |
|
SGEMM_INPUT_GS_BETA_NZ |
394 |
|
"vfmadd213ps %%ymm9, %%ymm3, %%ymm0 \n\t" |
395 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
396 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
397 |
|
" \n\t" |
398 |
|
" \n\t" |
399 |
|
SGEMM_INPUT_GS_BETA_NZ |
400 |
|
"vfmadd213ps %%ymm11, %%ymm3, %%ymm0 \n\t" |
401 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
402 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
403 |
|
" \n\t" |
404 |
|
" \n\t" |
405 |
|
SGEMM_INPUT_GS_BETA_NZ |
406 |
|
"vfmadd213ps %%ymm13, %%ymm3, %%ymm0 \n\t" |
407 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
408 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
409 |
|
" \n\t" |
410 |
|
" \n\t" |
411 |
|
SGEMM_INPUT_GS_BETA_NZ |
412 |
|
"vfmadd213ps %%ymm15, %%ymm3, %%ymm0 \n\t" |
413 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
414 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
415 |
|
" \n\t" |
416 |
|
" \n\t" |
417 |
|
" \n\t" |
418 |
|
"jmp .SDONE \n\t" // jump to end. |
419 |
|
" \n\t" |
420 |
|
" \n\t" |
421 |
|
" \n\t" |
422 |
|
".SCOLSTORED: \n\t" |
423 |
|
" \n\t" |
424 |
|
" \n\t" |
425 |
|
"vfmadd231ps (%%rcx), %%ymm3, %%ymm4 \n\t" |
426 |
|
"vmovups %%ymm4, (%%rcx) \n\t" |
427 |
|
"addq %%rdi, %%rcx \n\t" |
428 |
|
"vfmadd231ps (%%rdx), %%ymm3, %%ymm5 \n\t" |
429 |
|
"vmovups %%ymm5, (%%rdx) \n\t" |
430 |
|
"addq %%rdi, %%rdx \n\t" |
431 |
|
" \n\t" |
432 |
|
" \n\t" |
433 |
|
"vfmadd231ps (%%rcx), %%ymm3, %%ymm6 \n\t" |
434 |
|
"vmovups %%ymm6, (%%rcx) \n\t" |
435 |
|
"addq %%rdi, %%rcx \n\t" |
436 |
|
"vfmadd231ps (%%rdx), %%ymm3, %%ymm7 \n\t" |
437 |
|
"vmovups %%ymm7, (%%rdx) \n\t" |
438 |
|
"addq %%rdi, %%rdx \n\t" |
439 |
|
" \n\t" |
440 |
|
" \n\t" |
441 |
|
"vfmadd231ps (%%rcx), %%ymm3, %%ymm8 \n\t" |
442 |
|
"vmovups %%ymm8, (%%rcx) \n\t" |
443 |
|
"addq %%rdi, %%rcx \n\t" |
444 |
|
"vfmadd231ps (%%rdx), %%ymm3, %%ymm9 \n\t" |
445 |
|
"vmovups %%ymm9, (%%rdx) \n\t" |
446 |
|
"addq %%rdi, %%rdx \n\t" |
447 |
|
" \n\t" |
448 |
|
" \n\t" |
449 |
|
"vfmadd231ps (%%rcx), %%ymm3, %%ymm10 \n\t" |
450 |
|
"vmovups %%ymm10, (%%rcx) \n\t" |
451 |
|
"addq %%rdi, %%rcx \n\t" |
452 |
|
"vfmadd231ps (%%rdx), %%ymm3, %%ymm11 \n\t" |
453 |
|
"vmovups %%ymm11, (%%rdx) \n\t" |
454 |
|
"addq %%rdi, %%rdx \n\t" |
455 |
|
" \n\t" |
456 |
|
" \n\t" |
457 |
|
"vfmadd231ps (%%rcx), %%ymm3, %%ymm12 \n\t" |
458 |
|
"vmovups %%ymm12, (%%rcx) \n\t" |
459 |
|
"addq %%rdi, %%rcx \n\t" |
460 |
|
"vfmadd231ps (%%rdx), %%ymm3, %%ymm13 \n\t" |
461 |
|
"vmovups %%ymm13, (%%rdx) \n\t" |
462 |
|
"addq %%rdi, %%rdx \n\t" |
463 |
|
" \n\t" |
464 |
|
" \n\t" |
465 |
|
"vfmadd231ps (%%rcx), %%ymm3, %%ymm14 \n\t" |
466 |
|
"vmovups %%ymm14, (%%rcx) \n\t" |
467 |
|
//"addq %%rdi, %%rcx \n\t" |
468 |
|
"vfmadd231ps (%%rdx), %%ymm3, %%ymm15 \n\t" |
469 |
|
"vmovups %%ymm15, (%%rdx) \n\t" |
470 |
|
//"addq %%rdi, %%rdx \n\t" |
471 |
|
" \n\t" |
472 |
|
" \n\t" |
473 |
|
" \n\t" |
474 |
|
" \n\t" |
475 |
|
"jmp .SDONE \n\t" // jump to end. |
476 |
|
" \n\t" |
477 |
|
" \n\t" |
478 |
|
" \n\t" |
479 |
|
".SBETAZERO: \n\t" |
480 |
|
" \n\t" |
481 |
|
"cmpq $4, %%rsi \n\t" // set ZF if (4*rs_c) == 4. |
482 |
|
"jz .SCOLSTORBZ \n\t" // jump to column storage case |
483 |
|
" \n\t" |
484 |
|
" \n\t" |
485 |
|
" \n\t" |
486 |
|
".SGENSTORBZ: \n\t" |
487 |
|
" \n\t" |
488 |
|
" \n\t" |
489 |
|
"vmovaps %%ymm4, %%ymm0 \n\t" |
490 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
491 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
492 |
|
" \n\t" |
493 |
|
" \n\t" |
494 |
|
"vmovaps %%ymm6, %%ymm0 \n\t" |
495 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
496 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
497 |
|
" \n\t" |
498 |
|
" \n\t" |
499 |
|
"vmovaps %%ymm8, %%ymm0 \n\t" |
500 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
501 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
502 |
|
" \n\t" |
503 |
|
" \n\t" |
504 |
|
"vmovaps %%ymm10, %%ymm0 \n\t" |
505 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
506 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
507 |
|
" \n\t" |
508 |
|
" \n\t" |
509 |
|
"vmovaps %%ymm12, %%ymm0 \n\t" |
510 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
511 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
512 |
|
" \n\t" |
513 |
|
" \n\t" |
514 |
|
"vmovaps %%ymm14, %%ymm0 \n\t" |
515 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
516 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
517 |
|
" \n\t" |
518 |
|
" \n\t" |
519 |
|
"movq %%rdx, %%rcx \n\t" // rcx = c + 8*rs_c |
520 |
|
" \n\t" |
521 |
|
" \n\t" |
522 |
|
"vmovaps %%ymm5, %%ymm0 \n\t" |
523 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
524 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
525 |
|
" \n\t" |
526 |
|
" \n\t" |
527 |
|
"vmovaps %%ymm7, %%ymm0 \n\t" |
528 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
529 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
530 |
|
" \n\t" |
531 |
|
" \n\t" |
532 |
|
"vmovaps %%ymm9, %%ymm0 \n\t" |
533 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
534 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
535 |
|
" \n\t" |
536 |
|
" \n\t" |
537 |
|
"vmovaps %%ymm11, %%ymm0 \n\t" |
538 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
539 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
540 |
|
" \n\t" |
541 |
|
" \n\t" |
542 |
|
"vmovaps %%ymm13, %%ymm0 \n\t" |
543 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
544 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
545 |
|
" \n\t" |
546 |
|
" \n\t" |
547 |
|
"vmovaps %%ymm15, %%ymm0 \n\t" |
548 |
|
SGEMM_OUTPUT_GS_BETA_NZ |
549 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
550 |
|
" \n\t" |
551 |
|
" \n\t" |
552 |
|
" \n\t" |
553 |
|
"jmp .SDONE \n\t" // jump to end. |
554 |
|
" \n\t" |
555 |
|
" \n\t" |
556 |
|
" \n\t" |
557 |
|
".SCOLSTORBZ: \n\t" |
558 |
|
" \n\t" |
559 |
|
" \n\t" |
560 |
|
"vmovups %%ymm4, (%%rcx) \n\t" |
561 |
|
"addq %%rdi, %%rcx \n\t" |
562 |
|
"vmovups %%ymm5, (%%rdx) \n\t" |
563 |
|
"addq %%rdi, %%rdx \n\t" |
564 |
|
" \n\t" |
565 |
|
"vmovups %%ymm6, (%%rcx) \n\t" |
566 |
|
"addq %%rdi, %%rcx \n\t" |
567 |
|
"vmovups %%ymm7, (%%rdx) \n\t" |
568 |
|
"addq %%rdi, %%rdx \n\t" |
569 |
|
" \n\t" |
570 |
|
" \n\t" |
571 |
|
"vmovups %%ymm8, (%%rcx) \n\t" |
572 |
|
"addq %%rdi, %%rcx \n\t" |
573 |
|
"vmovups %%ymm9, (%%rdx) \n\t" |
574 |
|
"addq %%rdi, %%rdx \n\t" |
575 |
|
" \n\t" |
576 |
|
" \n\t" |
577 |
|
"vmovups %%ymm10, (%%rcx) \n\t" |
578 |
|
"addq %%rdi, %%rcx \n\t" |
579 |
|
"vmovups %%ymm11, (%%rdx) \n\t" |
580 |
|
"addq %%rdi, %%rdx \n\t" |
581 |
|
" \n\t" |
582 |
|
" \n\t" |
583 |
|
"vmovups %%ymm12, (%%rcx) \n\t" |
584 |
|
"addq %%rdi, %%rcx \n\t" |
585 |
|
"vmovups %%ymm13, (%%rdx) \n\t" |
586 |
|
"addq %%rdi, %%rdx \n\t" |
587 |
|
" \n\t" |
588 |
|
" \n\t" |
589 |
|
"vmovups %%ymm14, (%%rcx) \n\t" |
590 |
|
//"addq %%rdi, %%rcx \n\t" |
591 |
|
"vmovups %%ymm15, (%%rdx) \n\t" |
592 |
|
//"addq %%rdi, %%rdx \n\t" |
593 |
|
" \n\t" |
594 |
|
" \n\t" |
595 |
|
" \n\t" |
596 |
|
" \n\t" |
597 |
|
" \n\t" |
598 |
|
" \n\t" |
599 |
|
" \n\t" |
600 |
|
".SDONE: \n\t" |
601 |
|
" \n\t" |
602 |
|
"vzeroupper \n\t" |
603 |
|
" \n\t" |
604 |
|
|
605 |
|
: // output operands (none) |
606 |
|
: // input operands |
607 |
|
"m" (k_iter), // 0 |
608 |
|
"m" (k_left), // 1 |
609 |
|
"m" (a), // 2 |
610 |
|
"m" (b), // 3 |
611 |
|
"m" (alpha), // 4 |
612 |
|
"m" (beta), // 5 |
613 |
|
"m" (c), // 6 |
614 |
|
"m" (rs_c), // 7 |
615 |
|
"m" (cs_c)/*, // 8 |
616 |
|
"m" (b_next), // 9 |
617 |
|
"m" (a_next)*/ // 10 |
618 |
|
: // register clobber list |
619 |
|
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", |
620 |
|
"r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", |
621 |
|
"xmm0", "xmm1", "xmm2", "xmm3", |
622 |
|
"xmm4", "xmm5", "xmm6", "xmm7", |
623 |
|
"xmm8", "xmm9", "xmm10", "xmm11", |
624 |
|
"xmm12", "xmm13", "xmm14", "xmm15", |
625 |
|
"memory" |
626 |
|
); |
627 |
|
} |
628 |
|
|
629 |
|
#define DGEMM_INPUT_GS_BETA_NZ \ |
630 |
|
"vmovlpd (%%rcx ), %%xmm0, %%xmm0 \n\t" \ |
631 |
|
"vmovhpd (%%rcx,%%rsi,1), %%xmm0, %%xmm0 \n\t" \ |
632 |
|
"vmovlpd (%%rcx,%%rsi,2), %%xmm1, %%xmm1 \n\t" \ |
633 |
|
"vmovhpd (%%rcx,%%r13 ), %%xmm1, %%xmm1 \n\t" \ |
634 |
|
"vperm2f128 $0x20, %%ymm1, %%ymm0, %%ymm0 \n\t" /*\ |
635 |
|
"vmovlps (%%rcx,%%rsi,4), %%xmm2, %%xmm2 \n\t" \ |
636 |
|
"vmovhps (%%rcx,%%r15 ), %%xmm2, %%xmm2 \n\t" \ |
637 |
|
"vmovlps (%%rcx,%%r13,2), %%xmm1, %%xmm1 \n\t" \ |
638 |
|
"vmovhps (%%rcx,%%r10 ), %%xmm1, %%xmm1 \n\t" \ |
639 |
|
"vperm2f128 $0x20, %%ymm1, %%ymm2, %%ymm2 \n\t"*/ |
640 |
|
|
641 |
|
#define DGEMM_OUTPUT_GS_BETA_NZ \ |
642 |
|
"vextractf128 $1, %%ymm0, %%xmm1 \n\t" \ |
643 |
|
"vmovlpd %%xmm0, (%%rcx ) \n\t" \ |
644 |
|
"vmovhpd %%xmm0, (%%rcx,%%rsi ) \n\t" \ |
645 |
|
"vmovlpd %%xmm1, (%%rcx,%%rsi,2) \n\t" \ |
646 |
|
"vmovhpd %%xmm1, (%%rcx,%%r13 ) \n\t" /*\ |
647 |
|
"vextractf128 $1, %%ymm2, %%xmm1 \n\t" \ |
648 |
|
"vmovlpd %%xmm2, (%%rcx,%%rsi,4) \n\t" \ |
649 |
|
"vmovhpd %%xmm2, (%%rcx,%%r15 ) \n\t" \ |
650 |
|
"vmovlpd %%xmm1, (%%rcx,%%r13,2) \n\t" \ |
651 |
|
"vmovhpd %%xmm1, (%%rcx,%%r10 ) \n\t"*/ |
652 |
|
|
653 |
|
void bli_dgemm_asm_8x6 |
654 |
|
( |
655 |
|
dim_t k, |
656 |
|
double* restrict alpha, |
657 |
|
double* restrict a, |
658 |
|
double* restrict b, |
659 |
|
double* restrict beta, |
660 |
|
double* restrict c, inc_t rs_c, inc_t cs_c, |
661 |
|
//auxinfo_t* restrict data, |
662 |
|
//cntx_t* restrict cntx |
663 |
|
aux_s<double, double, double, double> *aux |
664 |
|
) |
665 |
|
{ |
666 |
|
//void* a_next = bli_auxinfo_next_a( data ); |
667 |
|
//void* b_next = bli_auxinfo_next_b( data ); |
668 |
|
|
669 |
|
dim_t k_iter = k / 4; |
670 |
|
dim_t k_left = k % 4; |
671 |
|
|
672 |
|
__asm__ volatile |
673 |
|
( |
674 |
|
" \n\t" |
675 |
|
"vzeroall \n\t" // zero all xmm/ymm registers. |
676 |
|
" \n\t" |
677 |
|
" \n\t" |
678 |
|
"movq %2, %%rax \n\t" // load address of a. |
679 |
|
"movq %3, %%rbx \n\t" // load address of b. |
680 |
|
//"movq %9, %%r15 \n\t" // load address of b_next. |
681 |
|
" \n\t" |
682 |
|
"addq $32 * 4, %%rax \n\t" |
683 |
|
" \n\t" // initialize loop by pre-loading |
684 |
|
"vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
685 |
|
"vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
686 |
|
" \n\t" |
687 |
|
"movq %6, %%rcx \n\t" // load address of c |
688 |
|
"movq %8, %%rdi \n\t" // load cs_c |
689 |
|
"leaq (,%%rdi,8), %%rdi \n\t" // cs_c *= sizeof(double) |
690 |
|
" \n\t" |
691 |
|
"leaq (%%rdi,%%rdi,2), %%r13 \n\t" // r13 = 3*cs_c; |
692 |
|
"leaq (%%rcx,%%r13,1), %%rdx \n\t" // rdx = c + 3*cs_c; |
693 |
|
"prefetcht0 7 * 8(%%rcx) \n\t" // prefetch c + 0*cs_c |
694 |
|
"prefetcht0 7 * 8(%%rcx,%%rdi) \n\t" // prefetch c + 1*cs_c |
695 |
|
"prefetcht0 7 * 8(%%rcx,%%rdi,2) \n\t" // prefetch c + 2*cs_c |
696 |
|
"prefetcht0 7 * 8(%%rdx) \n\t" // prefetch c + 3*cs_c |
697 |
|
"prefetcht0 7 * 8(%%rdx,%%rdi) \n\t" // prefetch c + 4*cs_c |
698 |
|
"prefetcht0 7 * 8(%%rdx,%%rdi,2) \n\t" // prefetch c + 5*cs_c |
699 |
|
" \n\t" |
700 |
|
" \n\t" |
701 |
|
" \n\t" |
702 |
|
" \n\t" |
703 |
|
"movq %0, %%rsi \n\t" // i = k_iter; |
704 |
|
"testq %%rsi, %%rsi \n\t" // check i via logical AND. |
705 |
|
"je .DCONSIDKLEFT \n\t" // if i == 0, jump to code that |
706 |
|
" \n\t" // contains the k_left loop. |
707 |
|
" \n\t" |
708 |
|
" \n\t" |
709 |
|
".DLOOPKITER: \n\t" // MAIN LOOP |
710 |
|
" \n\t" |
711 |
|
" \n\t" |
712 |
|
" \n\t" // iteration 0 |
713 |
|
"prefetcht0 64 * 8(%%rax) \n\t" |
714 |
|
" \n\t" |
715 |
|
"vbroadcastsd 0 * 8(%%rbx), %%ymm2 \n\t" |
716 |
|
"vbroadcastsd 1 * 8(%%rbx), %%ymm3 \n\t" |
717 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
718 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
719 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
720 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
721 |
|
" \n\t" |
722 |
|
"vbroadcastsd 2 * 8(%%rbx), %%ymm2 \n\t" |
723 |
|
"vbroadcastsd 3 * 8(%%rbx), %%ymm3 \n\t" |
724 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
725 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
726 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
727 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
728 |
|
" \n\t" |
729 |
|
"vbroadcastsd 4 * 8(%%rbx), %%ymm2 \n\t" |
730 |
|
"vbroadcastsd 5 * 8(%%rbx), %%ymm3 \n\t" |
731 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
732 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
733 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
734 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
735 |
|
" \n\t" |
736 |
|
"vmovaps -2 * 32(%%rax), %%ymm0 \n\t" |
737 |
|
"vmovaps -1 * 32(%%rax), %%ymm1 \n\t" |
738 |
|
" \n\t" |
739 |
|
" \n\t" // iteration 1 |
740 |
|
"vbroadcastsd 6 * 8(%%rbx), %%ymm2 \n\t" |
741 |
|
"vbroadcastsd 7 * 8(%%rbx), %%ymm3 \n\t" |
742 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
743 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
744 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
745 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
746 |
|
" \n\t" |
747 |
|
"vbroadcastsd 8 * 8(%%rbx), %%ymm2 \n\t" |
748 |
|
"vbroadcastsd 9 * 8(%%rbx), %%ymm3 \n\t" |
749 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
750 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
751 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
752 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
753 |
|
" \n\t" |
754 |
|
"vbroadcastsd 10 * 8(%%rbx), %%ymm2 \n\t" |
755 |
|
"vbroadcastsd 11 * 8(%%rbx), %%ymm3 \n\t" |
756 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
757 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
758 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
759 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
760 |
|
" \n\t" |
761 |
|
"vmovaps 0 * 32(%%rax), %%ymm0 \n\t" |
762 |
|
"vmovaps 1 * 32(%%rax), %%ymm1 \n\t" |
763 |
|
" \n\t" |
764 |
|
" \n\t" // iteration 2 |
765 |
|
"prefetcht0 76 * 8(%%rax) \n\t" |
766 |
|
" \n\t" |
767 |
|
"vbroadcastsd 12 * 8(%%rbx), %%ymm2 \n\t" |
768 |
|
"vbroadcastsd 13 * 8(%%rbx), %%ymm3 \n\t" |
769 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
770 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
771 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
772 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
773 |
|
" \n\t" |
774 |
|
"vbroadcastsd 14 * 8(%%rbx), %%ymm2 \n\t" |
775 |
|
"vbroadcastsd 15 * 8(%%rbx), %%ymm3 \n\t" |
776 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
777 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
778 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
779 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
780 |
|
" \n\t" |
781 |
|
"vbroadcastsd 16 * 8(%%rbx), %%ymm2 \n\t" |
782 |
|
"vbroadcastsd 17 * 8(%%rbx), %%ymm3 \n\t" |
783 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
784 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
785 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
786 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
787 |
|
" \n\t" |
788 |
|
"vmovaps 2 * 32(%%rax), %%ymm0 \n\t" |
789 |
|
"vmovaps 3 * 32(%%rax), %%ymm1 \n\t" |
790 |
|
" \n\t" |
791 |
|
" \n\t" // iteration 3 |
792 |
|
"vbroadcastsd 18 * 8(%%rbx), %%ymm2 \n\t" |
793 |
|
"vbroadcastsd 19 * 8(%%rbx), %%ymm3 \n\t" |
794 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
795 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
796 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
797 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
798 |
|
" \n\t" |
799 |
|
"vbroadcastsd 20 * 8(%%rbx), %%ymm2 \n\t" |
800 |
|
"vbroadcastsd 21 * 8(%%rbx), %%ymm3 \n\t" |
801 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
802 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
803 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
804 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
805 |
|
" \n\t" |
806 |
|
"vbroadcastsd 22 * 8(%%rbx), %%ymm2 \n\t" |
807 |
|
"vbroadcastsd 23 * 8(%%rbx), %%ymm3 \n\t" |
808 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
809 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
810 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
811 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
812 |
|
" \n\t" |
813 |
|
"addq $4 * 8 * 8, %%rax \n\t" // a += 4*8 (unroll x mr) |
814 |
|
"addq $4 * 6 * 8, %%rbx \n\t" // b += 4*6 (unroll x nr) |
815 |
|
" \n\t" |
816 |
|
"vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
817 |
|
"vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
818 |
|
" \n\t" |
819 |
|
" \n\t" |
820 |
|
"decq %%rsi \n\t" // i -= 1; |
821 |
|
"jne .DLOOPKITER \n\t" // iterate again if i != 0. |
822 |
|
" \n\t" |
823 |
|
" \n\t" |
824 |
|
" \n\t" |
825 |
|
" \n\t" |
826 |
|
" \n\t" |
827 |
|
" \n\t" |
828 |
|
".DCONSIDKLEFT: \n\t" |
829 |
|
" \n\t" |
830 |
|
"movq %1, %%rsi \n\t" // i = k_left; |
831 |
|
"testq %%rsi, %%rsi \n\t" // check i via logical AND. |
832 |
|
"je .DPOSTACCUM \n\t" // if i == 0, we're done; jump to end. |
833 |
|
" \n\t" // else, we prepare to enter k_left loop. |
834 |
|
" \n\t" |
835 |
|
" \n\t" |
836 |
|
".DLOOPKLEFT: \n\t" // EDGE LOOP |
837 |
|
" \n\t" |
838 |
|
"prefetcht0 64 * 8(%%rax) \n\t" |
839 |
|
" \n\t" |
840 |
|
"vbroadcastsd 0 * 8(%%rbx), %%ymm2 \n\t" |
841 |
|
"vbroadcastsd 1 * 8(%%rbx), %%ymm3 \n\t" |
842 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
843 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
844 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
845 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
846 |
|
" \n\t" |
847 |
|
"vbroadcastsd 2 * 8(%%rbx), %%ymm2 \n\t" |
848 |
|
"vbroadcastsd 3 * 8(%%rbx), %%ymm3 \n\t" |
849 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
850 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
851 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
852 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
853 |
|
" \n\t" |
854 |
|
"vbroadcastsd 4 * 8(%%rbx), %%ymm2 \n\t" |
855 |
|
"vbroadcastsd 5 * 8(%%rbx), %%ymm3 \n\t" |
856 |
|
"vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
857 |
|
"vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
858 |
|
"vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
859 |
|
"vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
860 |
|
" \n\t" |
861 |
|
"addq $1 * 8 * 8, %%rax \n\t" // a += 1*8 (unroll x mr) |
862 |
|
"addq $1 * 6 * 8, %%rbx \n\t" // b += 1*6 (unroll x nr) |
863 |
|
" \n\t" |
864 |
|
"vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
865 |
|
"vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
866 |
|
" \n\t" |
867 |
|
" \n\t" |
868 |
|
"decq %%rsi \n\t" // i -= 1; |
869 |
|
"jne .DLOOPKLEFT \n\t" // iterate again if i != 0. |
870 |
|
" \n\t" |
871 |
|
" \n\t" |
872 |
|
" \n\t" |
873 |
|
".DPOSTACCUM: \n\t" |
874 |
|
" \n\t" |
875 |
|
" \n\t" |
876 |
|
" \n\t" |
877 |
|
" \n\t" |
878 |
|
"movq %4, %%rax \n\t" // load address of alpha |
879 |
|
"movq %5, %%rbx \n\t" // load address of beta |
880 |
|
"vbroadcastsd (%%rax), %%ymm0 \n\t" // load alpha and duplicate |
881 |
|
"vbroadcastsd (%%rbx), %%ymm3 \n\t" // load beta and duplicate |
882 |
|
" \n\t" |
883 |
|
"vmulpd %%ymm0, %%ymm4, %%ymm4 \n\t" // scale by alpha |
884 |
|
"vmulpd %%ymm0, %%ymm5, %%ymm5 \n\t" |
885 |
|
"vmulpd %%ymm0, %%ymm6, %%ymm6 \n\t" |
886 |
|
"vmulpd %%ymm0, %%ymm7, %%ymm7 \n\t" |
887 |
|
"vmulpd %%ymm0, %%ymm8, %%ymm8 \n\t" |
888 |
|
"vmulpd %%ymm0, %%ymm9, %%ymm9 \n\t" |
889 |
|
"vmulpd %%ymm0, %%ymm10, %%ymm10 \n\t" |
890 |
|
"vmulpd %%ymm0, %%ymm11, %%ymm11 \n\t" |
891 |
|
"vmulpd %%ymm0, %%ymm12, %%ymm12 \n\t" |
892 |
|
"vmulpd %%ymm0, %%ymm13, %%ymm13 \n\t" |
893 |
|
"vmulpd %%ymm0, %%ymm14, %%ymm14 \n\t" |
894 |
|
"vmulpd %%ymm0, %%ymm15, %%ymm15 \n\t" |
895 |
|
" \n\t" |
896 |
|
" \n\t" |
897 |
|
" \n\t" |
898 |
|
" \n\t" |
899 |
|
" \n\t" |
900 |
|
" \n\t" |
901 |
|
"movq %7, %%rsi \n\t" // load rs_c |
902 |
|
"leaq (,%%rsi,8), %%rsi \n\t" // rsi = rs_c * sizeof(double) |
903 |
|
" \n\t" |
904 |
|
"leaq (%%rcx,%%rsi,4), %%rdx \n\t" // load address of c + 4*rs_c; |
905 |
|
" \n\t" |
906 |
|
"leaq (%%rsi,%%rsi,2), %%r13 \n\t" // r13 = 3*rs_c; |
907 |
|
//"leaq (%%rsi,%%rsi,4), %%r15 \n\t" // r15 = 5*rs_c; |
908 |
|
//"leaq (%%r13,%%rsi,4), %%r10 \n\t" // r10 = 7*rs_c; |
909 |
|
" \n\t" |
910 |
|
" \n\t" |
911 |
|
" \n\t" // now avoid loading C if beta == 0 |
912 |
|
" \n\t" |
913 |
|
"vxorpd %%ymm0, %%ymm0, %%ymm0 \n\t" // set ymm0 to zero. |
914 |
|
"vucomisd %%xmm0, %%xmm3 \n\t" // set ZF if beta == 0. |
915 |
|
"je .DBETAZERO \n\t" // if ZF = 1, jump to beta == 0 case |
916 |
|
" \n\t" |
917 |
|
" \n\t" |
918 |
|
"cmpq $8, %%rsi \n\t" // set ZF if (8*rs_c) == 8. |
919 |
|
"jz .DCOLSTORED \n\t" // jump to column storage case |
920 |
|
" \n\t" |
921 |
|
" \n\t" |
922 |
|
" \n\t" |
923 |
|
".DGENSTORED: \n\t" |
924 |
|
" \n\t" |
925 |
|
" \n\t" |
926 |
|
DGEMM_INPUT_GS_BETA_NZ |
927 |
|
"vfmadd213pd %%ymm4, %%ymm3, %%ymm0 \n\t" |
928 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
929 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
930 |
|
" \n\t" |
931 |
|
" \n\t" |
932 |
|
DGEMM_INPUT_GS_BETA_NZ |
933 |
|
"vfmadd213pd %%ymm6, %%ymm3, %%ymm0 \n\t" |
934 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
935 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
936 |
|
" \n\t" |
937 |
|
" \n\t" |
938 |
|
DGEMM_INPUT_GS_BETA_NZ |
939 |
|
"vfmadd213pd %%ymm8, %%ymm3, %%ymm0 \n\t" |
940 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
941 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
942 |
|
" \n\t" |
943 |
|
" \n\t" |
944 |
|
DGEMM_INPUT_GS_BETA_NZ |
945 |
|
"vfmadd213pd %%ymm10, %%ymm3, %%ymm0 \n\t" |
946 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
947 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
948 |
|
" \n\t" |
949 |
|
" \n\t" |
950 |
|
DGEMM_INPUT_GS_BETA_NZ |
951 |
|
"vfmadd213pd %%ymm12, %%ymm3, %%ymm0 \n\t" |
952 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
953 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
954 |
|
" \n\t" |
955 |
|
" \n\t" |
956 |
|
DGEMM_INPUT_GS_BETA_NZ |
957 |
|
"vfmadd213pd %%ymm14, %%ymm3, %%ymm0 \n\t" |
958 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
959 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
960 |
|
" \n\t" |
961 |
|
" \n\t" |
962 |
|
"movq %%rdx, %%rcx \n\t" // rcx = c + 4*rs_c |
963 |
|
" \n\t" |
964 |
|
" \n\t" |
965 |
|
DGEMM_INPUT_GS_BETA_NZ |
966 |
|
"vfmadd213pd %%ymm5, %%ymm3, %%ymm0 \n\t" |
967 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
968 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
969 |
|
" \n\t" |
970 |
|
" \n\t" |
971 |
|
DGEMM_INPUT_GS_BETA_NZ |
972 |
|
"vfmadd213pd %%ymm7, %%ymm3, %%ymm0 \n\t" |
973 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
974 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
975 |
|
" \n\t" |
976 |
|
" \n\t" |
977 |
|
DGEMM_INPUT_GS_BETA_NZ |
978 |
|
"vfmadd213pd %%ymm9, %%ymm3, %%ymm0 \n\t" |
979 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
980 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
981 |
|
" \n\t" |
982 |
|
" \n\t" |
983 |
|
DGEMM_INPUT_GS_BETA_NZ |
984 |
|
"vfmadd213pd %%ymm11, %%ymm3, %%ymm0 \n\t" |
985 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
986 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
987 |
|
" \n\t" |
988 |
|
" \n\t" |
989 |
|
DGEMM_INPUT_GS_BETA_NZ |
990 |
|
"vfmadd213pd %%ymm13, %%ymm3, %%ymm0 \n\t" |
991 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
992 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
993 |
|
" \n\t" |
994 |
|
" \n\t" |
995 |
|
DGEMM_INPUT_GS_BETA_NZ |
996 |
|
"vfmadd213pd %%ymm15, %%ymm3, %%ymm0 \n\t" |
997 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
998 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
999 |
|
" \n\t" |
1000 |
|
" \n\t" |
1001 |
|
" \n\t" |
1002 |
|
"jmp .DDONE \n\t" // jump to end. |
1003 |
|
" \n\t" |
1004 |
|
" \n\t" |
1005 |
|
" \n\t" |
1006 |
|
".DCOLSTORED: \n\t" |
1007 |
|
" \n\t" |
1008 |
|
" \n\t" |
1009 |
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm4 \n\t" |
1010 |
|
"vmovups %%ymm4, (%%rcx) \n\t" |
1011 |
|
"addq %%rdi, %%rcx \n\t" |
1012 |
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm5 \n\t" |
1013 |
|
"vmovups %%ymm5, (%%rdx) \n\t" |
1014 |
|
"addq %%rdi, %%rdx \n\t" |
1015 |
|
" \n\t" |
1016 |
|
" \n\t" |
1017 |
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm6 \n\t" |
1018 |
|
"vmovups %%ymm6, (%%rcx) \n\t" |
1019 |
|
"addq %%rdi, %%rcx \n\t" |
1020 |
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm7 \n\t" |
1021 |
|
"vmovups %%ymm7, (%%rdx) \n\t" |
1022 |
|
"addq %%rdi, %%rdx \n\t" |
1023 |
|
" \n\t" |
1024 |
|
" \n\t" |
1025 |
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm8 \n\t" |
1026 |
|
"vmovups %%ymm8, (%%rcx) \n\t" |
1027 |
|
"addq %%rdi, %%rcx \n\t" |
1028 |
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm9 \n\t" |
1029 |
|
"vmovups %%ymm9, (%%rdx) \n\t" |
1030 |
|
"addq %%rdi, %%rdx \n\t" |
1031 |
|
" \n\t" |
1032 |
|
" \n\t" |
1033 |
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm10 \n\t" |
1034 |
|
"vmovups %%ymm10, (%%rcx) \n\t" |
1035 |
|
"addq %%rdi, %%rcx \n\t" |
1036 |
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm11 \n\t" |
1037 |
|
"vmovups %%ymm11, (%%rdx) \n\t" |
1038 |
|
"addq %%rdi, %%rdx \n\t" |
1039 |
|
" \n\t" |
1040 |
|
" \n\t" |
1041 |
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm12 \n\t" |
1042 |
|
"vmovups %%ymm12, (%%rcx) \n\t" |
1043 |
|
"addq %%rdi, %%rcx \n\t" |
1044 |
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm13 \n\t" |
1045 |
|
"vmovups %%ymm13, (%%rdx) \n\t" |
1046 |
|
"addq %%rdi, %%rdx \n\t" |
1047 |
|
" \n\t" |
1048 |
|
" \n\t" |
1049 |
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm14 \n\t" |
1050 |
|
"vmovups %%ymm14, (%%rcx) \n\t" |
1051 |
|
//"addq %%rdi, %%rcx \n\t" |
1052 |
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm15 \n\t" |
1053 |
|
"vmovups %%ymm15, (%%rdx) \n\t" |
1054 |
|
//"addq %%rdi, %%rdx \n\t" |
1055 |
|
" \n\t" |
1056 |
|
" \n\t" |
1057 |
|
" \n\t" |
1058 |
|
"jmp .DDONE \n\t" // jump to end. |
1059 |
|
" \n\t" |
1060 |
|
" \n\t" |
1061 |
|
" \n\t" |
1062 |
|
".DBETAZERO: \n\t" |
1063 |
|
" \n\t" |
1064 |
|
"cmpq $8, %%rsi \n\t" // set ZF if (8*rs_c) == 8. |
1065 |
|
"jz .DCOLSTORBZ \n\t" // jump to column storage case |
1066 |
|
" \n\t" |
1067 |
|
" \n\t" |
1068 |
|
" \n\t" |
1069 |
|
".DGENSTORBZ: \n\t" |
1070 |
|
" \n\t" |
1071 |
|
" \n\t" |
1072 |
|
"vmovaps %%ymm4, %%ymm0 \n\t" |
1073 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1074 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1075 |
|
" \n\t" |
1076 |
|
" \n\t" |
1077 |
|
"vmovaps %%ymm6, %%ymm0 \n\t" |
1078 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1079 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1080 |
|
" \n\t" |
1081 |
|
" \n\t" |
1082 |
|
"vmovaps %%ymm8, %%ymm0 \n\t" |
1083 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1084 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1085 |
|
" \n\t" |
1086 |
|
" \n\t" |
1087 |
|
"vmovaps %%ymm10, %%ymm0 \n\t" |
1088 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1089 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1090 |
|
" \n\t" |
1091 |
|
" \n\t" |
1092 |
|
"vmovaps %%ymm12, %%ymm0 \n\t" |
1093 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1094 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1095 |
|
" \n\t" |
1096 |
|
" \n\t" |
1097 |
|
"vmovaps %%ymm14, %%ymm0 \n\t" |
1098 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1099 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1100 |
|
" \n\t" |
1101 |
|
" \n\t" |
1102 |
|
"movq %%rdx, %%rcx \n\t" // rcx = c + 4*rs_c |
1103 |
|
" \n\t" |
1104 |
|
" \n\t" |
1105 |
|
"vmovaps %%ymm5, %%ymm0 \n\t" |
1106 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1107 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1108 |
|
" \n\t" |
1109 |
|
" \n\t" |
1110 |
|
"vmovaps %%ymm7, %%ymm0 \n\t" |
1111 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1112 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1113 |
|
" \n\t" |
1114 |
|
" \n\t" |
1115 |
|
"vmovaps %%ymm9, %%ymm0 \n\t" |
1116 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1117 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1118 |
|
" \n\t" |
1119 |
|
" \n\t" |
1120 |
|
"vmovaps %%ymm11, %%ymm0 \n\t" |
1121 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1122 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1123 |
|
" \n\t" |
1124 |
|
" \n\t" |
1125 |
|
"vmovaps %%ymm13, %%ymm0 \n\t" |
1126 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1127 |
|
"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1128 |
|
" \n\t" |
1129 |
|
" \n\t" |
1130 |
|
"vmovaps %%ymm15, %%ymm0 \n\t" |
1131 |
|
DGEMM_OUTPUT_GS_BETA_NZ |
1132 |
|
//"addq %%rdi, %%rcx \n\t" // c += cs_c; |
1133 |
|
" \n\t" |
1134 |
|
" \n\t" |
1135 |
|
" \n\t" |
1136 |
|
"jmp .DDONE \n\t" // jump to end. |
1137 |
|
" \n\t" |
1138 |
|
" \n\t" |
1139 |
|
" \n\t" |
1140 |
|
".DCOLSTORBZ: \n\t" |
1141 |
|
" \n\t" |
1142 |
|
" \n\t" |
1143 |
|
"vmovups %%ymm4, (%%rcx) \n\t" |
1144 |
|
"addq %%rdi, %%rcx \n\t" |
1145 |
|
"vmovups %%ymm5, (%%rdx) \n\t" |
1146 |
|
"addq %%rdi, %%rdx \n\t" |
1147 |
|
" \n\t" |
1148 |
|
"vmovups %%ymm6, (%%rcx) \n\t" |
1149 |
|
"addq %%rdi, %%rcx \n\t" |
1150 |
|
"vmovups %%ymm7, (%%rdx) \n\t" |
1151 |
|
"addq %%rdi, %%rdx \n\t" |
1152 |
|
" \n\t" |
1153 |
|
" \n\t" |
1154 |
|
"vmovups %%ymm8, (%%rcx) \n\t" |
1155 |
|
"addq %%rdi, %%rcx \n\t" |
1156 |
|
"vmovups %%ymm9, (%%rdx) \n\t" |
1157 |
|
"addq %%rdi, %%rdx \n\t" |
1158 |
|
" \n\t" |
1159 |
|
" \n\t" |
1160 |
|
"vmovups %%ymm10, (%%rcx) \n\t" |
1161 |
|
"addq %%rdi, %%rcx \n\t" |
1162 |
|
"vmovups %%ymm11, (%%rdx) \n\t" |
1163 |
|
"addq %%rdi, %%rdx \n\t" |
1164 |
|
" \n\t" |
1165 |
|
" \n\t" |
1166 |
|
"vmovups %%ymm12, (%%rcx) \n\t" |
1167 |
|
"addq %%rdi, %%rcx \n\t" |
1168 |
|
"vmovups %%ymm13, (%%rdx) \n\t" |
1169 |
|
"addq %%rdi, %%rdx \n\t" |
1170 |
|
" \n\t" |
1171 |
|
" \n\t" |
1172 |
|
"vmovups %%ymm14, (%%rcx) \n\t" |
1173 |
|
//"addq %%rdi, %%rcx \n\t" |
1174 |
|
"vmovups %%ymm15, (%%rdx) \n\t" |
1175 |
|
//"addq %%rdi, %%rdx \n\t" |
1176 |
|
" \n\t" |
1177 |
|
" \n\t" |
1178 |
|
" \n\t" |
1179 |
|
" \n\t" |
1180 |
|
" \n\t" |
1181 |
|
" \n\t" |
1182 |
|
" \n\t" |
1183 |
|
".DDONE: \n\t" |
1184 |
|
" \n\t" |
1185 |
|
"vzeroupper \n\t" |
1186 |
|
" \n\t" |
1187 |
|
|
1188 |
|
: // output operands (none) |
1189 |
|
: // input operands |
1190 |
|
"m" (k_iter), // 0 |
1191 |
|
"m" (k_left), // 1 |
1192 |
|
"m" (a), // 2 |
1193 |
|
"m" (b), // 3 |
1194 |
|
"m" (alpha), // 4 |
1195 |
|
"m" (beta), // 5 |
1196 |
|
"m" (c), // 6 |
1197 |
|
"m" (rs_c), // 7 |
1198 |
|
"m" (cs_c)/*, // 8 |
1199 |
|
"m" (b_next), // 9 |
1200 |
|
"m" (a_next)*/ // 10 |
1201 |
|
: // register clobber list |
1202 |
|
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", |
1203 |
|
"r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", |
1204 |
|
"xmm0", "xmm1", "xmm2", "xmm3", |
1205 |
|
"xmm4", "xmm5", "xmm6", "xmm7", |
1206 |
|
"xmm8", "xmm9", "xmm10", "xmm11", |
1207 |
|
"xmm12", "xmm13", "xmm14", "xmm15", |
1208 |
|
"memory" |
1209 |
|
); |
1210 |
|
} |
1211 |
|
|
1212 |
|
|
1213 |
|
|
1214 |
|
|
1215 |
|
// assumes beta.r, beta.i have been broadcast into ymm1, ymm2. |
1216 |
|
// outputs to ymm0 |
1217 |
|
#define CGEMM_INPUT_SCALE_GS_BETA_NZ \ |
1218 |
|
"vmovlpd (%%rcx ), %%xmm0, %%xmm0 \n\t" \ |
1219 |
|
"vmovhpd (%%rcx,%%rsi,1), %%xmm0, %%xmm0 \n\t" \ |
1220 |
|
"vmovlpd (%%rcx,%%rsi,2), %%xmm3, %%xmm3 \n\t" \ |
1221 |
|
"vmovhpd (%%rcx,%%r13 ), %%xmm3, %%xmm3 \n\t" \ |
1222 |
|
"vinsertf128 $1, %%xmm3, %%ymm0, %%ymm0 \n\t" \ |
1223 |
|
"vpermilps $0xb1, %%ymm0, %%ymm3 \n\t" \ |
1224 |
|
"vmulps %%ymm1, %%ymm0, %%ymm0 \n\t" \ |
1225 |
|
"vmulps %%ymm2, %%ymm3, %%ymm3 \n\t" \ |
1226 |
|
"vaddsubps %%ymm3, %%ymm0, %%ymm0 \n\t" |
1227 |
|
|
1228 |
|
// assumes values to output are in ymm0 |
1229 |
|
#define CGEMM_OUTPUT_GS \ |
1230 |
|
"vextractf128 $1, %%ymm0, %%xmm3 \n\t" \ |
1231 |
|
"vmovlpd %%xmm0, (%%rcx ) \n\t" \ |
1232 |
|
"vmovhpd %%xmm0, (%%rcx,%%rsi,1) \n\t" \ |
1233 |
|
"vmovlpd %%xmm3, (%%rcx,%%rsi,2) \n\t" \ |
1234 |
|
"vmovhpd %%xmm3, (%%rcx,%%r13 ) \n\t" |
1235 |
|
|
1236 |
|
#define CGEMM_INPUT_SCALE_CS_BETA_NZ \ |
1237 |
|
"vmovups (%%rcx), %%ymm0 \n\t" \ |
1238 |
|
"vpermilps $0xb1, %%ymm0, %%ymm3 \n\t" \ |
1239 |
|
"vmulps %%ymm1, %%ymm0, %%ymm0 \n\t" \ |
1240 |
|
"vmulps %%ymm2, %%ymm3, %%ymm3 \n\t" \ |
1241 |
|
"vaddsubps %%ymm3, %%ymm0, %%ymm0 \n\t" |
1242 |
|
|
1243 |
|
#define CGEMM_OUTPUT_CS \ |
1244 |
|
"vmovups %%ymm0, (%%rcx) \n\t" \ |
1245 |
|
|
1246 |
|
//void bli_cgemm_asm_8x3 |
1247 |
|
// ( |
1248 |
|
// dim_t k, |
1249 |
|
// scomplex* restrict alpha, |
1250 |
|
// scomplex* restrict a, |
1251 |
|
// scomplex* restrict b, |
1252 |
|
// scomplex* restrict beta, |
1253 |
|
// scomplex* restrict c, inc_t rs_c, inc_t cs_c, |
1254 |
|
// auxinfo_t* restrict data, |
1255 |
|
// cntx_t* restrict cntx |
1256 |
|
// ) |
1257 |
|
//{ |
1258 |
|
// //void* a_next = bli_auxinfo_next_a( data ); |
1259 |
|
// //void* b_next = bli_auxinfo_next_b( data ); |
1260 |
|
// |
1261 |
|
// uint64_t k_iter = k / 4; |
1262 |
|
// uint64_t k_left = k % 4; |
1263 |
|
// |
1264 |
|
// __asm__ volatile |
1265 |
|
// ( |
1266 |
|
// " \n\t" |
1267 |
|
// "vzeroall \n\t" // zero all xmm/ymm registers. |
1268 |
|
// " \n\t" |
1269 |
|
// " \n\t" |
1270 |
|
// "movq %2, %%rax \n\t" // load address of a. |
1271 |
|
// "movq %3, %%rbx \n\t" // load address of b. |
1272 |
|
// //"movq %9, %%r15 \n\t" // load address of b_next. |
1273 |
|
// " \n\t" |
1274 |
|
// "addq $32 * 4, %%rax \n\t" |
1275 |
|
// " \n\t" // initialize loop by pre-loading |
1276 |
|
// "vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
1277 |
|
// "vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
1278 |
|
// " \n\t" |
1279 |
|
// "movq %6, %%rcx \n\t" // load address of c |
1280 |
|
// "movq %8, %%rdi \n\t" // load cs_c |
1281 |
|
// "leaq (,%%rdi,8), %%rdi \n\t" // cs_c *= sizeof(scomplex) |
1282 |
|
// " \n\t" |
1283 |
|
// "leaq (%%rcx,%%rdi,1), %%r11 \n\t" // r11 = c + 1*cs_c; |
1284 |
|
// "leaq (%%rcx,%%rdi,2), %%r12 \n\t" // r12 = c + 2*cs_c; |
1285 |
|
// " \n\t" |
1286 |
|
// "prefetcht0 7 * 8(%%rcx) \n\t" // prefetch c + 0*cs_c |
1287 |
|
// "prefetcht0 7 * 8(%%r11) \n\t" // prefetch c + 1*cs_c |
1288 |
|
// "prefetcht0 7 * 8(%%r12) \n\t" // prefetch c + 2*cs_c |
1289 |
|
// " \n\t" |
1290 |
|
// " \n\t" |
1291 |
|
// " \n\t" |
1292 |
|
// " \n\t" |
1293 |
|
// "movq %0, %%rsi \n\t" // i = k_iter; |
1294 |
|
// "testq %%rsi, %%rsi \n\t" // check i via logical AND. |
1295 |
|
// "je .CCONSIDKLEFT \n\t" // if i == 0, jump to code that |
1296 |
|
// " \n\t" // contains the k_left loop. |
1297 |
|
// " \n\t" |
1298 |
|
// " \n\t" |
1299 |
|
// ".CLOOPKITER: \n\t" // MAIN LOOP |
1300 |
|
// " \n\t" |
1301 |
|
// " \n\t" |
1302 |
|
// " \n\t" // iteration 0 |
1303 |
|
// "prefetcht0 32 * 8(%%rax) \n\t" |
1304 |
|
// " \n\t" |
1305 |
|
// "vbroadcastss 0 * 4(%%rbx), %%ymm2 \n\t" |
1306 |
|
// "vbroadcastss 1 * 4(%%rbx), %%ymm3 \n\t" |
1307 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
1308 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
1309 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
1310 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
1311 |
|
// " \n\t" |
1312 |
|
// "vbroadcastss 2 * 4(%%rbx), %%ymm2 \n\t" |
1313 |
|
// "vbroadcastss 3 * 4(%%rbx), %%ymm3 \n\t" |
1314 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
1315 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
1316 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
1317 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
1318 |
|
// " \n\t" |
1319 |
|
// "vbroadcastss 4 * 4(%%rbx), %%ymm2 \n\t" |
1320 |
|
// "vbroadcastss 5 * 4(%%rbx), %%ymm3 \n\t" |
1321 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
1322 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
1323 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
1324 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
1325 |
|
// " \n\t" |
1326 |
|
// "vmovaps -2 * 32(%%rax), %%ymm0 \n\t" |
1327 |
|
// "vmovaps -1 * 32(%%rax), %%ymm1 \n\t" |
1328 |
|
// " \n\t" |
1329 |
|
// " \n\t" // iteration 1 |
1330 |
|
// "vbroadcastss 6 * 4(%%rbx), %%ymm2 \n\t" |
1331 |
|
// "vbroadcastss 7 * 4(%%rbx), %%ymm3 \n\t" |
1332 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
1333 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
1334 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
1335 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
1336 |
|
// " \n\t" |
1337 |
|
// "vbroadcastss 8 * 4(%%rbx), %%ymm2 \n\t" |
1338 |
|
// "vbroadcastss 9 * 4(%%rbx), %%ymm3 \n\t" |
1339 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
1340 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
1341 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
1342 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
1343 |
|
// " \n\t" |
1344 |
|
// "vbroadcastss 10 * 4(%%rbx), %%ymm2 \n\t" |
1345 |
|
// "vbroadcastss 11 * 4(%%rbx), %%ymm3 \n\t" |
1346 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
1347 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
1348 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
1349 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
1350 |
|
// " \n\t" |
1351 |
|
// "vmovaps 0 * 32(%%rax), %%ymm0 \n\t" |
1352 |
|
// "vmovaps 1 * 32(%%rax), %%ymm1 \n\t" |
1353 |
|
// " \n\t" |
1354 |
|
// " \n\t" // iteration 2 |
1355 |
|
// "prefetcht0 38 * 8(%%rax) \n\t" |
1356 |
|
// " \n\t" |
1357 |
|
// "vbroadcastss 12 * 4(%%rbx), %%ymm2 \n\t" |
1358 |
|
// "vbroadcastss 13 * 4(%%rbx), %%ymm3 \n\t" |
1359 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
1360 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
1361 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
1362 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
1363 |
|
// " \n\t" |
1364 |
|
// "vbroadcastss 14 * 4(%%rbx), %%ymm2 \n\t" |
1365 |
|
// "vbroadcastss 15 * 4(%%rbx), %%ymm3 \n\t" |
1366 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
1367 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
1368 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
1369 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
1370 |
|
// " \n\t" |
1371 |
|
// "vbroadcastss 16 * 4(%%rbx), %%ymm2 \n\t" |
1372 |
|
// "vbroadcastss 17 * 4(%%rbx), %%ymm3 \n\t" |
1373 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
1374 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
1375 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
1376 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
1377 |
|
// " \n\t" |
1378 |
|
// "vmovaps 2 * 32(%%rax), %%ymm0 \n\t" |
1379 |
|
// "vmovaps 3 * 32(%%rax), %%ymm1 \n\t" |
1380 |
|
// " \n\t" |
1381 |
|
// " \n\t" // iteration 3 |
1382 |
|
// "vbroadcastss 18 * 4(%%rbx), %%ymm2 \n\t" |
1383 |
|
// "vbroadcastss 19 * 4(%%rbx), %%ymm3 \n\t" |
1384 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
1385 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
1386 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
1387 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
1388 |
|
// " \n\t" |
1389 |
|
// "vbroadcastss 20 * 4(%%rbx), %%ymm2 \n\t" |
1390 |
|
// "vbroadcastss 21 * 4(%%rbx), %%ymm3 \n\t" |
1391 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
1392 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
1393 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
1394 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
1395 |
|
// " \n\t" |
1396 |
|
// "vbroadcastss 22 * 4(%%rbx), %%ymm2 \n\t" |
1397 |
|
// "vbroadcastss 23 * 4(%%rbx), %%ymm3 \n\t" |
1398 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
1399 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
1400 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
1401 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
1402 |
|
// " \n\t" |
1403 |
|
// "addq $4 * 8 * 8, %%rax \n\t" // a += 4*8 (unroll x mr) |
1404 |
|
// "addq $4 * 3 * 8, %%rbx \n\t" // b += 4*3 (unroll x nr) |
1405 |
|
// " \n\t" |
1406 |
|
// "vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
1407 |
|
// "vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
1408 |
|
// " \n\t" |
1409 |
|
// " \n\t" |
1410 |
|
// "decq %%rsi \n\t" // i -= 1; |
1411 |
|
// "jne .CLOOPKITER \n\t" // iterate again if i != 0. |
1412 |
|
// " \n\t" |
1413 |
|
// " \n\t" |
1414 |
|
// " \n\t" |
1415 |
|
// " \n\t" |
1416 |
|
// " \n\t" |
1417 |
|
// " \n\t" |
1418 |
|
// ".CCONSIDKLEFT: \n\t" |
1419 |
|
// " \n\t" |
1420 |
|
// "movq %1, %%rsi \n\t" // i = k_left; |
1421 |
|
// "testq %%rsi, %%rsi \n\t" // check i via logical AND. |
1422 |
|
// "je .CPOSTACCUM \n\t" // if i == 0, we're done; jump to end. |
1423 |
|
// " \n\t" // else, we prepare to enter k_left loop. |
1424 |
|
// " \n\t" |
1425 |
|
// " \n\t" |
1426 |
|
// ".CLOOPKLEFT: \n\t" // EDGE LOOP |
1427 |
|
// " \n\t" |
1428 |
|
// "prefetcht0 32 * 8(%%rax) \n\t" |
1429 |
|
// " \n\t" |
1430 |
|
// "vbroadcastss 0 * 4(%%rbx), %%ymm2 \n\t" |
1431 |
|
// "vbroadcastss 1 * 4(%%rbx), %%ymm3 \n\t" |
1432 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm4 \n\t" |
1433 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm5 \n\t" |
1434 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm6 \n\t" |
1435 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm7 \n\t" |
1436 |
|
// " \n\t" |
1437 |
|
// "vbroadcastss 2 * 4(%%rbx), %%ymm2 \n\t" |
1438 |
|
// "vbroadcastss 3 * 4(%%rbx), %%ymm3 \n\t" |
1439 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm8 \n\t" |
1440 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm9 \n\t" |
1441 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm10 \n\t" |
1442 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm11 \n\t" |
1443 |
|
// " \n\t" |
1444 |
|
// "vbroadcastss 4 * 4(%%rbx), %%ymm2 \n\t" |
1445 |
|
// "vbroadcastss 5 * 4(%%rbx), %%ymm3 \n\t" |
1446 |
|
// "vfmadd231ps %%ymm0, %%ymm2, %%ymm12 \n\t" |
1447 |
|
// "vfmadd231ps %%ymm1, %%ymm2, %%ymm13 \n\t" |
1448 |
|
// "vfmadd231ps %%ymm0, %%ymm3, %%ymm14 \n\t" |
1449 |
|
// "vfmadd231ps %%ymm1, %%ymm3, %%ymm15 \n\t" |
1450 |
|
// " \n\t" |
1451 |
|
// "addq $1 * 8 * 8, %%rax \n\t" // a += 1*8 (unroll x mr) |
1452 |
|
// "addq $1 * 3 * 8, %%rbx \n\t" // b += 1*3 (unroll x nr) |
1453 |
|
// " \n\t" |
1454 |
|
// "vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
1455 |
|
// "vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
1456 |
|
// " \n\t" |
1457 |
|
// " \n\t" |
1458 |
|
// "decq %%rsi \n\t" // i -= 1; |
1459 |
|
// "jne .CLOOPKLEFT \n\t" // iterate again if i != 0. |
1460 |
|
// " \n\t" |
1461 |
|
// " \n\t" |
1462 |
|
// " \n\t" |
1463 |
|
// ".CPOSTACCUM: \n\t" |
1464 |
|
// " \n\t" |
1465 |
|
// " \n\t" |
1466 |
|
// " \n\t" // permute even and odd elements |
1467 |
|
// " \n\t" // of ymm6/7, ymm10/11, ymm/14/15 |
1468 |
|
// "vpermilps $0xb1, %%ymm6, %%ymm6 \n\t" |
1469 |
|
// "vpermilps $0xb1, %%ymm7, %%ymm7 \n\t" |
1470 |
|
// "vpermilps $0xb1, %%ymm10, %%ymm10 \n\t" |
1471 |
|
// "vpermilps $0xb1, %%ymm11, %%ymm11 \n\t" |
1472 |
|
// "vpermilps $0xb1, %%ymm14, %%ymm14 \n\t" |
1473 |
|
// "vpermilps $0xb1, %%ymm15, %%ymm15 \n\t" |
1474 |
|
// " \n\t" |
1475 |
|
// " \n\t" |
1476 |
|
// " \n\t" // subtract/add even/odd elements |
1477 |
|
// "vaddsubps %%ymm6, %%ymm4, %%ymm4 \n\t" |
1478 |
|
// "vaddsubps %%ymm7, %%ymm5, %%ymm5 \n\t" |
1479 |
|
// " \n\t" |
1480 |
|
// "vaddsubps %%ymm10, %%ymm8, %%ymm8 \n\t" |
1481 |
|
// "vaddsubps %%ymm11, %%ymm9, %%ymm9 \n\t" |
1482 |
|
// " \n\t" |
1483 |
|
// "vaddsubps %%ymm14, %%ymm12, %%ymm12 \n\t" |
1484 |
|
// "vaddsubps %%ymm15, %%ymm13, %%ymm13 \n\t" |
1485 |
|
// " \n\t" |
1486 |
|
// " \n\t" |
1487 |
|
// " \n\t" |
1488 |
|
// " \n\t" |
1489 |
|
// "movq %4, %%rax \n\t" // load address of alpha |
1490 |
|
// "vbroadcastss (%%rax), %%ymm0 \n\t" // load alpha_r and duplicate |
1491 |
|
// "vbroadcastss 4(%%rax), %%ymm1 \n\t" // load alpha_i and duplicate |
1492 |
|
// " \n\t" |
1493 |
|
// " \n\t" |
1494 |
|
// "vpermilps $0xb1, %%ymm4, %%ymm3 \n\t" |
1495 |
|
// "vmulps %%ymm0, %%ymm4, %%ymm4 \n\t" |
1496 |
|
// "vmulps %%ymm1, %%ymm3, %%ymm3 \n\t" |
1497 |
|
// "vaddsubps %%ymm3, %%ymm4, %%ymm4 \n\t" |
1498 |
|
// " \n\t" |
1499 |
|
// "vpermilps $0xb1, %%ymm5, %%ymm3 \n\t" |
1500 |
|
// "vmulps %%ymm0, %%ymm5, %%ymm5 \n\t" |
1501 |
|
// "vmulps %%ymm1, %%ymm3, %%ymm3 \n\t" |
1502 |
|
// "vaddsubps %%ymm3, %%ymm5, %%ymm5 \n\t" |
1503 |
|
// " \n\t" |
1504 |
|
// " \n\t" |
1505 |
|
// "vpermilps $0xb1, %%ymm8, %%ymm3 \n\t" |
1506 |
|
// "vmulps %%ymm0, %%ymm8, %%ymm8 \n\t" |
1507 |
|
// "vmulps %%ymm1, %%ymm3, %%ymm3 \n\t" |
1508 |
|
// "vaddsubps %%ymm3, %%ymm8, %%ymm8 \n\t" |
1509 |
|
// " \n\t" |
1510 |
|
// "vpermilps $0xb1, %%ymm9, %%ymm3 \n\t" |
1511 |
|
// "vmulps %%ymm0, %%ymm9, %%ymm9 \n\t" |
1512 |
|
// "vmulps %%ymm1, %%ymm3, %%ymm3 \n\t" |
1513 |
|
// "vaddsubps %%ymm3, %%ymm9, %%ymm9 \n\t" |
1514 |
|
// " \n\t" |
1515 |
|
// " \n\t" |
1516 |
|
// "vpermilps $0xb1, %%ymm12, %%ymm3 \n\t" |
1517 |
|
// "vmulps %%ymm0, %%ymm12, %%ymm12 \n\t" |
1518 |
|
// "vmulps %%ymm1, %%ymm3, %%ymm3 \n\t" |
1519 |
|
// "vaddsubps %%ymm3, %%ymm12, %%ymm12 \n\t" |
1520 |
|
// " \n\t" |
1521 |
|
// "vpermilps $0xb1, %%ymm13, %%ymm3 \n\t" |
1522 |
|
// "vmulps %%ymm0, %%ymm13, %%ymm13 \n\t" |
1523 |
|
// "vmulps %%ymm1, %%ymm3, %%ymm3 \n\t" |
1524 |
|
// "vaddsubps %%ymm3, %%ymm13, %%ymm13 \n\t" |
1525 |
|
// " \n\t" |
1526 |
|
// " \n\t" |
1527 |
|
// " \n\t" |
1528 |
|
// " \n\t" |
1529 |
|
// " \n\t" |
1530 |
|
// "movq %5, %%rbx \n\t" // load address of beta |
1531 |
|
// "vbroadcastss (%%rbx), %%ymm1 \n\t" // load beta_r and duplicate |
1532 |
|
// "vbroadcastss 4(%%rbx), %%ymm2 \n\t" // load beta_i and duplicate |
1533 |
|
// " \n\t" |
1534 |
|
// " \n\t" |
1535 |
|
// " \n\t" |
1536 |
|
// " \n\t" |
1537 |
|
// "movq %7, %%rsi \n\t" // load rs_c |
1538 |
|
// "leaq (,%%rsi,8), %%rsi \n\t" // rsi = rs_c * sizeof(scomplex) |
1539 |
|
// "leaq (,%%rsi,4), %%rdx \n\t" // rdx = 4*rs_c; |
1540 |
|
// "leaq (%%rsi,%%rsi,2), %%r13 \n\t" // r13 = 3*rs_c; |
1541 |
|
// " \n\t" |
1542 |
|
// " \n\t" |
1543 |
|
// " \n\t" |
1544 |
|
// " \n\t" // now avoid loading C if beta == 0 |
1545 |
|
// "vxorps %%ymm0, %%ymm0, %%ymm0 \n\t" // set ymm0 to zero. |
1546 |
|
// "vucomiss %%xmm0, %%xmm1 \n\t" // set ZF if beta_r == 0. |
1547 |
|
// "sete %%r8b \n\t" // r8b = ( ZF == 1 ? 1 : 0 ); |
1548 |
|
// "vucomiss %%xmm0, %%xmm2 \n\t" // set ZF if beta_i == 0. |
1549 |
|
// "sete %%r9b \n\t" // r9b = ( ZF == 1 ? 1 : 0 ); |
1550 |
|
// "andb %%r8b, %%r9b \n\t" // set ZF if r8b & r9b == 1. |
1551 |
|
// "jne .CBETAZERO \n\t" // if ZF = 1, jump to beta == 0 case |
1552 |
|
// " \n\t" |
1553 |
|
// " \n\t" |
1554 |
|
// "cmpq $8, %%rsi \n\t" // set ZF if (8*cs_c) == 8. |
1555 |
|
// "jz .CCOLSTORED \n\t" // jump to row storage case |
1556 |
|
// " \n\t" |
1557 |
|
// " \n\t" |
1558 |
|
// " \n\t" |
1559 |
|
// ".CGENSTORED: \n\t" |
1560 |
|
// " \n\t" |
1561 |
|
// " \n\t" |
1562 |
|
// CGEMM_INPUT_SCALE_GS_BETA_NZ |
1563 |
|
// "vaddps %%ymm4, %%ymm0, %%ymm0 \n\t" |
1564 |
|
// CGEMM_OUTPUT_GS |
1565 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1566 |
|
// " \n\t" |
1567 |
|
// " \n\t" |
1568 |
|
// CGEMM_INPUT_SCALE_GS_BETA_NZ |
1569 |
|
// "vaddps %%ymm5, %%ymm0, %%ymm0 \n\t" |
1570 |
|
// CGEMM_OUTPUT_GS |
1571 |
|
// "movq %%r11, %%rcx \n\t" // rcx = c + 1*cs_c |
1572 |
|
// " \n\t" |
1573 |
|
// " \n\t" |
1574 |
|
// " \n\t" |
1575 |
|
// CGEMM_INPUT_SCALE_GS_BETA_NZ |
1576 |
|
// "vaddps %%ymm8, %%ymm0, %%ymm0 \n\t" |
1577 |
|
// CGEMM_OUTPUT_GS |
1578 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1579 |
|
// " \n\t" |
1580 |
|
// " \n\t" |
1581 |
|
// CGEMM_INPUT_SCALE_GS_BETA_NZ |
1582 |
|
// "vaddps %%ymm9, %%ymm0, %%ymm0 \n\t" |
1583 |
|
// CGEMM_OUTPUT_GS |
1584 |
|
// "movq %%r12, %%rcx \n\t" // rcx = c + 2*cs_c |
1585 |
|
// " \n\t" |
1586 |
|
// " \n\t" |
1587 |
|
// " \n\t" |
1588 |
|
// CGEMM_INPUT_SCALE_GS_BETA_NZ |
1589 |
|
// "vaddps %%ymm12, %%ymm0, %%ymm0 \n\t" |
1590 |
|
// CGEMM_OUTPUT_GS |
1591 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1592 |
|
// " \n\t" |
1593 |
|
// " \n\t" |
1594 |
|
// CGEMM_INPUT_SCALE_GS_BETA_NZ |
1595 |
|
// "vaddps %%ymm13, %%ymm0, %%ymm0 \n\t" |
1596 |
|
// CGEMM_OUTPUT_GS |
1597 |
|
// " \n\t" |
1598 |
|
// " \n\t" |
1599 |
|
// " \n\t" |
1600 |
|
// "jmp .CDONE \n\t" // jump to end. |
1601 |
|
// " \n\t" |
1602 |
|
// " \n\t" |
1603 |
|
// " \n\t" |
1604 |
|
// ".CCOLSTORED: \n\t" |
1605 |
|
// " \n\t" |
1606 |
|
// " \n\t" |
1607 |
|
// CGEMM_INPUT_SCALE_CS_BETA_NZ |
1608 |
|
// "vaddps %%ymm4, %%ymm0, %%ymm0 \n\t" |
1609 |
|
// CGEMM_OUTPUT_CS |
1610 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1611 |
|
// " \n\t" |
1612 |
|
// " \n\t" |
1613 |
|
// CGEMM_INPUT_SCALE_CS_BETA_NZ |
1614 |
|
// "vaddps %%ymm5, %%ymm0, %%ymm0 \n\t" |
1615 |
|
// CGEMM_OUTPUT_CS |
1616 |
|
// "movq %%r11, %%rcx \n\t" // rcx = c + 1*cs_c |
1617 |
|
// " \n\t" |
1618 |
|
// " \n\t" |
1619 |
|
// " \n\t" |
1620 |
|
// CGEMM_INPUT_SCALE_CS_BETA_NZ |
1621 |
|
// "vaddps %%ymm8, %%ymm0, %%ymm0 \n\t" |
1622 |
|
// CGEMM_OUTPUT_CS |
1623 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1624 |
|
// " \n\t" |
1625 |
|
// " \n\t" |
1626 |
|
// CGEMM_INPUT_SCALE_CS_BETA_NZ |
1627 |
|
// "vaddps %%ymm9, %%ymm0, %%ymm0 \n\t" |
1628 |
|
// CGEMM_OUTPUT_CS |
1629 |
|
// "movq %%r12, %%rcx \n\t" // rcx = c + 2*cs_c |
1630 |
|
// " \n\t" |
1631 |
|
// " \n\t" |
1632 |
|
// " \n\t" |
1633 |
|
// CGEMM_INPUT_SCALE_CS_BETA_NZ |
1634 |
|
// "vaddps %%ymm12, %%ymm0, %%ymm0 \n\t" |
1635 |
|
// CGEMM_OUTPUT_CS |
1636 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1637 |
|
// " \n\t" |
1638 |
|
// " \n\t" |
1639 |
|
// CGEMM_INPUT_SCALE_CS_BETA_NZ |
1640 |
|
// "vaddps %%ymm13, %%ymm0, %%ymm0 \n\t" |
1641 |
|
// CGEMM_OUTPUT_CS |
1642 |
|
// " \n\t" |
1643 |
|
// " \n\t" |
1644 |
|
// " \n\t" |
1645 |
|
// "jmp .CDONE \n\t" // jump to end. |
1646 |
|
// " \n\t" |
1647 |
|
// " \n\t" |
1648 |
|
// " \n\t" |
1649 |
|
// ".CBETAZERO: \n\t" |
1650 |
|
// " \n\t" |
1651 |
|
// "cmpq $8, %%rsi \n\t" // set ZF if (8*rs_c) == 8. |
1652 |
|
// "jz .CCOLSTORBZ \n\t" // jump to row storage case |
1653 |
|
// " \n\t" |
1654 |
|
// " \n\t" |
1655 |
|
// " \n\t" |
1656 |
|
// ".CGENSTORBZ: \n\t" |
1657 |
|
// " \n\t" |
1658 |
|
// " \n\t" |
1659 |
|
// "vmovaps %%ymm4, %%ymm0 \n\t" |
1660 |
|
// CGEMM_OUTPUT_GS |
1661 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1662 |
|
// " \n\t" |
1663 |
|
// " \n\t" |
1664 |
|
// "vmovaps %%ymm5, %%ymm0 \n\t" |
1665 |
|
// CGEMM_OUTPUT_GS |
1666 |
|
// "movq %%r11, %%rcx \n\t" // rcx = c + 1*cs_c |
1667 |
|
// " \n\t" |
1668 |
|
// " \n\t" |
1669 |
|
// " \n\t" |
1670 |
|
// "vmovaps %%ymm8, %%ymm0 \n\t" |
1671 |
|
// CGEMM_OUTPUT_GS |
1672 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1673 |
|
// " \n\t" |
1674 |
|
// " \n\t" |
1675 |
|
// "vmovaps %%ymm9, %%ymm0 \n\t" |
1676 |
|
// CGEMM_OUTPUT_GS |
1677 |
|
// "movq %%r12, %%rcx \n\t" // rcx = c + 2*cs_c |
1678 |
|
// " \n\t" |
1679 |
|
// " \n\t" |
1680 |
|
// " \n\t" |
1681 |
|
// "vmovaps %%ymm12, %%ymm0 \n\t" |
1682 |
|
// CGEMM_OUTPUT_GS |
1683 |
|
// "addq %%rdx, %%rcx \n\t" // c += 4*rs_c; |
1684 |
|
// " \n\t" |
1685 |
|
// " \n\t" |
1686 |
|
// "vmovaps %%ymm13, %%ymm0 \n\t" |
1687 |
|
// CGEMM_OUTPUT_GS |
1688 |
|
// " \n\t" |
1689 |
|
// " \n\t" |
1690 |
|
// " \n\t" |
1691 |
|
// "jmp .CDONE \n\t" // jump to end. |
1692 |
|
// " \n\t" |
1693 |
|
// " \n\t" |
1694 |
|
// " \n\t" |
1695 |
|
// ".CCOLSTORBZ: \n\t" |
1696 |
|
// " \n\t" |
1697 |
|
// " \n\t" |
1698 |
|
// "vmovups %%ymm4, (%%rcx) \n\t" |
1699 |
|
// "vmovups %%ymm5, (%%rcx,%%rdx,1) \n\t" |
1700 |
|
// " \n\t" |
1701 |
|
// "vmovups %%ymm8, (%%r11) \n\t" |
1702 |
|
// "vmovups %%ymm9, (%%r11,%%rdx,1) \n\t" |
1703 |
|
// " \n\t" |
1704 |
|
// "vmovups %%ymm12, (%%r12) \n\t" |
1705 |
|
// "vmovups %%ymm13, (%%r12,%%rdx,1) \n\t" |
1706 |
|
// " \n\t" |
1707 |
|
// " \n\t" |
1708 |
|
// " \n\t" |
1709 |
|
// " \n\t" |
1710 |
|
// " \n\t" |
1711 |
|
// " \n\t" |
1712 |
|
// ".CDONE: \n\t" |
1713 |
|
// " \n\t" |
1714 |
|
// "vzeroupper \n\t" |
1715 |
|
// " \n\t" |
1716 |
|
// |
1717 |
|
// : // output operands (none) |
1718 |
|
// : // input operands |
1719 |
|
// "m" (k_iter), // 0 |
1720 |
|
// "m" (k_left), // 1 |
1721 |
|
// "m" (a), // 2 |
1722 |
|
// "m" (b), // 3 |
1723 |
|
// "m" (alpha), // 4 |
1724 |
|
// "m" (beta), // 5 |
1725 |
|
// "m" (c), // 6 |
1726 |
|
// "m" (rs_c), // 7 |
1727 |
|
// "m" (cs_c)/*, // 8 |
1728 |
|
// "m" (b_next), // 9 |
1729 |
|
// "m" (a_next)*/ // 10 |
1730 |
|
// : // register clobber list |
1731 |
|
// "rax", "rbx", "rcx", "rdx", "rsi", "rdi", |
1732 |
|
// "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", |
1733 |
|
// "xmm0", "xmm1", "xmm2", "xmm3", |
1734 |
|
// "xmm4", "xmm5", "xmm6", "xmm7", |
1735 |
|
// "xmm8", "xmm9", "xmm10", "xmm11", |
1736 |
|
// "xmm12", "xmm13", "xmm14", "xmm15", |
1737 |
|
// "memory" |
1738 |
|
// ); |
1739 |
|
//} |
1740 |
|
// |
1741 |
|
// |
1742 |
|
// |
1743 |
|
// |
1744 |
|
//// assumes beta.r, beta.i have been broadcast into ymm1, ymm2. |
1745 |
|
//// outputs to ymm0 |
1746 |
|
//#define ZGEMM_INPUT_SCALE_GS_BETA_NZ \ |
1747 |
|
// "vmovupd (%%rcx), %%xmm0 \n\t" \ |
1748 |
|
// "vmovupd (%%rcx,%%rsi), %%xmm3 \n\t" \ |
1749 |
|
// "vinsertf128 $1, %%xmm3, %%ymm0, %%ymm0 \n\t" \ |
1750 |
|
// "vpermilpd $0x5, %%ymm0, %%ymm3 \n\t" \ |
1751 |
|
// "vmulpd %%ymm1, %%ymm0, %%ymm0 \n\t" \ |
1752 |
|
// "vmulpd %%ymm2, %%ymm3, %%ymm3 \n\t" \ |
1753 |
|
// "vaddsubpd %%ymm3, %%ymm0, %%ymm0 \n\t" |
1754 |
|
// |
1755 |
|
//// assumes values to output are in ymm0 |
1756 |
|
//#define ZGEMM_OUTPUT_GS \ |
1757 |
|
// "vextractf128 $1, %%ymm0, %%xmm3 \n\t" \ |
1758 |
|
// "vmovupd %%xmm0, (%%rcx) \n\t" \ |
1759 |
|
// "vmovupd %%xmm3, (%%rcx,%%rsi ) \n\t" \ |
1760 |
|
// |
1761 |
|
//#define ZGEMM_INPUT_SCALE_CS_BETA_NZ \ |
1762 |
|
// "vmovups (%%rcx), %%ymm0 \n\t" \ |
1763 |
|
// "vpermilpd $0x5, %%ymm0, %%ymm3 \n\t" \ |
1764 |
|
// "vmulpd %%ymm1, %%ymm0, %%ymm0 \n\t" \ |
1765 |
|
// "vmulpd %%ymm2, %%ymm3, %%ymm3 \n\t" \ |
1766 |
|
// "vaddsubpd %%ymm3, %%ymm0, %%ymm0 \n\t" |
1767 |
|
// |
1768 |
|
//#define ZGEMM_OUTPUT_CS \ |
1769 |
|
// "vmovupd %%ymm0, (%%rcx) \n\t" \ |
1770 |
|
// |
1771 |
|
//void bli_zgemm_asm_4x3 |
1772 |
|
// ( |
1773 |
|
// dim_t k, |
1774 |
|
// dcomplex* restrict alpha, |
1775 |
|
// dcomplex* restrict a, |
1776 |
|
// dcomplex* restrict b, |
1777 |
|
// dcomplex* restrict beta, |
1778 |
|
// dcomplex* restrict c, inc_t rs_c, inc_t cs_c, |
1779 |
|
// auxinfo_t* restrict data, |
1780 |
|
// cntx_t* restrict cntx |
1781 |
|
// ) |
1782 |
|
//{ |
1783 |
|
// //void* a_next = bli_auxinfo_next_a( data ); |
1784 |
|
// //void* b_next = bli_auxinfo_next_b( data ); |
1785 |
|
// |
1786 |
|
// uint64_t k_iter = k / 4; |
1787 |
|
// uint64_t k_left = k % 4; |
1788 |
|
// |
1789 |
|
// //uint64_t alpha_is_unit = bli_zeq1( *alpha ); |
1790 |
|
// |
1791 |
|
// |
1792 |
|
// __asm__ volatile |
1793 |
|
// ( |
1794 |
|
// " \n\t" |
1795 |
|
// "vzeroall \n\t" // zero all xmm/ymm registers. |
1796 |
|
// " \n\t" |
1797 |
|
// " \n\t" |
1798 |
|
// "movq %2, %%rax \n\t" // load address of a. |
1799 |
|
// "movq %3, %%rbx \n\t" // load address of b. |
1800 |
|
// //"movq %9, %%r15 \n\t" // load address of b_next. |
1801 |
|
// " \n\t" |
1802 |
|
// "addq $32 * 4, %%rax \n\t" |
1803 |
|
// " \n\t" // initialize loop by pre-loading |
1804 |
|
// "vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
1805 |
|
// "vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
1806 |
|
// " \n\t" |
1807 |
|
// "movq %6, %%rcx \n\t" // load address of c |
1808 |
|
// "movq %8, %%rdi \n\t" // load cs_c |
1809 |
|
// "leaq (,%%rdi,8), %%rdi \n\t" // cs_c *= sizeof(dcomplex) |
1810 |
|
// "leaq (,%%rdi,2), %%rdi \n\t" |
1811 |
|
// " \n\t" |
1812 |
|
// "leaq (%%rcx,%%rdi,1), %%r11 \n\t" // r11 = c + 1*cs_c; |
1813 |
|
// "leaq (%%rcx,%%rdi,2), %%r12 \n\t" // r12 = c + 2*cs_c; |
1814 |
|
// " \n\t" |
1815 |
|
// "prefetcht0 7 * 8(%%rcx) \n\t" // prefetch c + 0*cs_c |
1816 |
|
// "prefetcht0 7 * 8(%%r11) \n\t" // prefetch c + 1*cs_c |
1817 |
|
// "prefetcht0 7 * 8(%%r12) \n\t" // prefetch c + 2*cs_c |
1818 |
|
// " \n\t" |
1819 |
|
// " \n\t" |
1820 |
|
// " \n\t" |
1821 |
|
// " \n\t" |
1822 |
|
// "movq %0, %%rsi \n\t" // i = k_iter; |
1823 |
|
// "testq %%rsi, %%rsi \n\t" // check i via logical AND. |
1824 |
|
// "je .ZCONSIDKLEFT \n\t" // if i == 0, jump to code that |
1825 |
|
// " \n\t" // contains the k_left loop. |
1826 |
|
// " \n\t" |
1827 |
|
// " \n\t" |
1828 |
|
// ".ZLOOPKITER: \n\t" // MAIN LOOP |
1829 |
|
// " \n\t" |
1830 |
|
// " \n\t" |
1831 |
|
// " \n\t" // iteration 0 |
1832 |
|
// "prefetcht0 32 * 16(%%rax) \n\t" |
1833 |
|
// " \n\t" |
1834 |
|
// "vbroadcastsd 0 * 8(%%rbx), %%ymm2 \n\t" |
1835 |
|
// "vbroadcastsd 1 * 8(%%rbx), %%ymm3 \n\t" |
1836 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
1837 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
1838 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
1839 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
1840 |
|
// " \n\t" |
1841 |
|
// "vbroadcastsd 2 * 8(%%rbx), %%ymm2 \n\t" |
1842 |
|
// "vbroadcastsd 3 * 8(%%rbx), %%ymm3 \n\t" |
1843 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
1844 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
1845 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
1846 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
1847 |
|
// " \n\t" |
1848 |
|
// "vbroadcastsd 4 * 8(%%rbx), %%ymm2 \n\t" |
1849 |
|
// "vbroadcastsd 5 * 8(%%rbx), %%ymm3 \n\t" |
1850 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
1851 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
1852 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
1853 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
1854 |
|
// " \n\t" |
1855 |
|
// "vmovaps -2 * 32(%%rax), %%ymm0 \n\t" |
1856 |
|
// "vmovaps -1 * 32(%%rax), %%ymm1 \n\t" |
1857 |
|
// " \n\t" |
1858 |
|
// " \n\t" // iteration 1 |
1859 |
|
// "vbroadcastsd 6 * 8(%%rbx), %%ymm2 \n\t" |
1860 |
|
// "vbroadcastsd 7 * 8(%%rbx), %%ymm3 \n\t" |
1861 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
1862 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
1863 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
1864 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
1865 |
|
// " \n\t" |
1866 |
|
// "vbroadcastsd 8 * 8(%%rbx), %%ymm2 \n\t" |
1867 |
|
// "vbroadcastsd 9 * 8(%%rbx), %%ymm3 \n\t" |
1868 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
1869 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
1870 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
1871 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
1872 |
|
// " \n\t" |
1873 |
|
// "vbroadcastsd 10 * 8(%%rbx), %%ymm2 \n\t" |
1874 |
|
// "vbroadcastsd 11 * 8(%%rbx), %%ymm3 \n\t" |
1875 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
1876 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
1877 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
1878 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
1879 |
|
// " \n\t" |
1880 |
|
// "vmovaps 0 * 32(%%rax), %%ymm0 \n\t" |
1881 |
|
// "vmovaps 1 * 32(%%rax), %%ymm1 \n\t" |
1882 |
|
// " \n\t" |
1883 |
|
// " \n\t" // iteration 2 |
1884 |
|
// "prefetcht0 38 * 16(%%rax) \n\t" |
1885 |
|
// " \n\t" |
1886 |
|
// "vbroadcastsd 12 * 8(%%rbx), %%ymm2 \n\t" |
1887 |
|
// "vbroadcastsd 13 * 8(%%rbx), %%ymm3 \n\t" |
1888 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
1889 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
1890 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
1891 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
1892 |
|
// " \n\t" |
1893 |
|
// "vbroadcastsd 14 * 8(%%rbx), %%ymm2 \n\t" |
1894 |
|
// "vbroadcastsd 15 * 8(%%rbx), %%ymm3 \n\t" |
1895 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
1896 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
1897 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
1898 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
1899 |
|
// " \n\t" |
1900 |
|
// "vbroadcastsd 16 * 8(%%rbx), %%ymm2 \n\t" |
1901 |
|
// "vbroadcastsd 17 * 8(%%rbx), %%ymm3 \n\t" |
1902 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
1903 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
1904 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
1905 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
1906 |
|
// " \n\t" |
1907 |
|
// "vmovaps 2 * 32(%%rax), %%ymm0 \n\t" |
1908 |
|
// "vmovaps 3 * 32(%%rax), %%ymm1 \n\t" |
1909 |
|
// " \n\t" |
1910 |
|
// " \n\t" // iteration 3 |
1911 |
|
// "vbroadcastsd 18 * 8(%%rbx), %%ymm2 \n\t" |
1912 |
|
// "vbroadcastsd 19 * 8(%%rbx), %%ymm3 \n\t" |
1913 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
1914 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
1915 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
1916 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
1917 |
|
// " \n\t" |
1918 |
|
// "vbroadcastsd 20 * 8(%%rbx), %%ymm2 \n\t" |
1919 |
|
// "vbroadcastsd 21 * 8(%%rbx), %%ymm3 \n\t" |
1920 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
1921 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
1922 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
1923 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
1924 |
|
// " \n\t" |
1925 |
|
// "vbroadcastsd 22 * 8(%%rbx), %%ymm2 \n\t" |
1926 |
|
// "vbroadcastsd 23 * 8(%%rbx), %%ymm3 \n\t" |
1927 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
1928 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
1929 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
1930 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
1931 |
|
// " \n\t" |
1932 |
|
// "addq $4 * 4 * 16, %%rax \n\t" // a += 4*4 (unroll x mr) |
1933 |
|
// "addq $4 * 3 * 16, %%rbx \n\t" // b += 4*3 (unroll x nr) |
1934 |
|
// " \n\t" |
1935 |
|
// "vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
1936 |
|
// "vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
1937 |
|
// " \n\t" |
1938 |
|
// " \n\t" |
1939 |
|
// "decq %%rsi \n\t" // i -= 1; |
1940 |
|
// "jne .ZLOOPKITER \n\t" // iterate again if i != 0. |
1941 |
|
// " \n\t" |
1942 |
|
// " \n\t" |
1943 |
|
// " \n\t" |
1944 |
|
// " \n\t" |
1945 |
|
// " \n\t" |
1946 |
|
// " \n\t" |
1947 |
|
// ".ZCONSIDKLEFT: \n\t" |
1948 |
|
// " \n\t" |
1949 |
|
// "movq %1, %%rsi \n\t" // i = k_left; |
1950 |
|
// "testq %%rsi, %%rsi \n\t" // check i via logical AND. |
1951 |
|
// "je .ZPOSTACCUM \n\t" // if i == 0, we're done; jump to end. |
1952 |
|
// " \n\t" // else, we prepare to enter k_left loop. |
1953 |
|
// " \n\t" |
1954 |
|
// " \n\t" |
1955 |
|
// ".ZLOOPKLEFT: \n\t" // EDGE LOOP |
1956 |
|
// " \n\t" |
1957 |
|
// "prefetcht0 32 * 16(%%rax) \n\t" |
1958 |
|
// " \n\t" |
1959 |
|
// "vbroadcastsd 0 * 8(%%rbx), %%ymm2 \n\t" |
1960 |
|
// "vbroadcastsd 1 * 8(%%rbx), %%ymm3 \n\t" |
1961 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm4 \n\t" |
1962 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm5 \n\t" |
1963 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm6 \n\t" |
1964 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm7 \n\t" |
1965 |
|
// " \n\t" |
1966 |
|
// "vbroadcastsd 2 * 8(%%rbx), %%ymm2 \n\t" |
1967 |
|
// "vbroadcastsd 3 * 8(%%rbx), %%ymm3 \n\t" |
1968 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm8 \n\t" |
1969 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm9 \n\t" |
1970 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm10 \n\t" |
1971 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm11 \n\t" |
1972 |
|
// " \n\t" |
1973 |
|
// "vbroadcastsd 4 * 8(%%rbx), %%ymm2 \n\t" |
1974 |
|
// "vbroadcastsd 5 * 8(%%rbx), %%ymm3 \n\t" |
1975 |
|
// "vfmadd231pd %%ymm0, %%ymm2, %%ymm12 \n\t" |
1976 |
|
// "vfmadd231pd %%ymm1, %%ymm2, %%ymm13 \n\t" |
1977 |
|
// "vfmadd231pd %%ymm0, %%ymm3, %%ymm14 \n\t" |
1978 |
|
// "vfmadd231pd %%ymm1, %%ymm3, %%ymm15 \n\t" |
1979 |
|
// " \n\t" |
1980 |
|
// "addq $1 * 4 * 16, %%rax \n\t" // a += 1*4 (unroll x mr) |
1981 |
|
// "addq $1 * 3 * 16, %%rbx \n\t" // b += 1*3 (unroll x nr) |
1982 |
|
// " \n\t" |
1983 |
|
// "vmovaps -4 * 32(%%rax), %%ymm0 \n\t" |
1984 |
|
// "vmovaps -3 * 32(%%rax), %%ymm1 \n\t" |
1985 |
|
// " \n\t" |
1986 |
|
// " \n\t" |
1987 |
|
// "decq %%rsi \n\t" // i -= 1; |
1988 |
|
// "jne .ZLOOPKLEFT \n\t" // iterate again if i != 0. |
1989 |
|
// " \n\t" |
1990 |
|
// " \n\t" |
1991 |
|
// " \n\t" |
1992 |
|
// ".ZPOSTACCUM: \n\t" |
1993 |
|
// " \n\t" |
1994 |
|
// " \n\t" // permute even and odd elements |
1995 |
|
// " \n\t" // of ymm6/7, ymm10/11, ymm/14/15 |
1996 |
|
// "vpermilpd $0x5, %%ymm6, %%ymm6 \n\t" |
1997 |
|
// "vpermilpd $0x5, %%ymm7, %%ymm7 \n\t" |
1998 |
|
// "vpermilpd $0x5, %%ymm10, %%ymm10 \n\t" |
1999 |
|
// "vpermilpd $0x5, %%ymm11, %%ymm11 \n\t" |
2000 |
|
// "vpermilpd $0x5, %%ymm14, %%ymm14 \n\t" |
2001 |
|
// "vpermilpd $0x5, %%ymm15, %%ymm15 \n\t" |
2002 |
|
// " \n\t" |
2003 |
|
// " \n\t" |
2004 |
|
// " \n\t" // subtract/add even/odd elements |
2005 |
|
// "vaddsubpd %%ymm6, %%ymm4, %%ymm4 \n\t" |
2006 |
|
// "vaddsubpd %%ymm7, %%ymm5, %%ymm5 \n\t" |
2007 |
|
// " \n\t" |
2008 |
|
// "vaddsubpd %%ymm10, %%ymm8, %%ymm8 \n\t" |
2009 |
|
// "vaddsubpd %%ymm11, %%ymm9, %%ymm9 \n\t" |
2010 |
|
// " \n\t" |
2011 |
|
// "vaddsubpd %%ymm14, %%ymm12, %%ymm12 \n\t" |
2012 |
|
// "vaddsubpd %%ymm15, %%ymm13, %%ymm13 \n\t" |
2013 |
|
// " \n\t" |
2014 |
|
// " \n\t" |
2015 |
|
// " \n\t" |
2016 |
|
// " \n\t" |
2017 |
|
// "movq %4, %%rax \n\t" // load address of alpha |
2018 |
|
// "vbroadcastsd (%%rax), %%ymm0 \n\t" // load alpha_r and duplicate |
2019 |
|
// "vbroadcastsd 8(%%rax), %%ymm1 \n\t" // load alpha_i and duplicate |
2020 |
|
// " \n\t" |
2021 |
|
// " \n\t" |
2022 |
|
// "vpermilpd $0x5, %%ymm4, %%ymm3 \n\t" |
2023 |
|
// "vmulpd %%ymm0, %%ymm4, %%ymm4 \n\t" |
2024 |
|
// "vmulpd %%ymm1, %%ymm3, %%ymm3 \n\t" |
2025 |
|
// "vaddsubpd %%ymm3, %%ymm4, %%ymm4 \n\t" |
2026 |
|
// " \n\t" |
2027 |
|
// "vpermilpd $0x5, %%ymm5, %%ymm3 \n\t" |
2028 |
|
// "vmulpd %%ymm0, %%ymm5, %%ymm5 \n\t" |
2029 |
|
// "vmulpd %%ymm1, %%ymm3, %%ymm3 \n\t" |
2030 |
|
// "vaddsubpd %%ymm3, %%ymm5, %%ymm5 \n\t" |
2031 |
|
// " \n\t" |
2032 |
|
// " \n\t" |
2033 |
|
// "vpermilpd $0x5, %%ymm8, %%ymm3 \n\t" |
2034 |
|
// "vmulpd %%ymm0, %%ymm8, %%ymm8 \n\t" |
2035 |
|
// "vmulpd %%ymm1, %%ymm3, %%ymm3 \n\t" |
2036 |
|
// "vaddsubpd %%ymm3, %%ymm8, %%ymm8 \n\t" |
2037 |
|
// " \n\t" |
2038 |
|
// "vpermilpd $0x5, %%ymm9, %%ymm3 \n\t" |
2039 |
|
// "vmulpd %%ymm0, %%ymm9, %%ymm9 \n\t" |
2040 |
|
// "vmulpd %%ymm1, %%ymm3, %%ymm3 \n\t" |
2041 |
|
// "vaddsubpd %%ymm3, %%ymm9, %%ymm9 \n\t" |
2042 |
|
// " \n\t" |
2043 |
|
// " \n\t" |
2044 |
|
// "vpermilpd $0x5, %%ymm12, %%ymm3 \n\t" |
2045 |
|
// "vmulpd %%ymm0, %%ymm12, %%ymm12 \n\t" |
2046 |
|
// "vmulpd %%ymm1, %%ymm3, %%ymm3 \n\t" |
2047 |
|
// "vaddsubpd %%ymm3, %%ymm12, %%ymm12 \n\t" |
2048 |
|
// " \n\t" |
2049 |
|
// "vpermilpd $0x5, %%ymm13, %%ymm3 \n\t" |
2050 |
|
// "vmulpd %%ymm0, %%ymm13, %%ymm13 \n\t" |
2051 |
|
// "vmulpd %%ymm1, %%ymm3, %%ymm3 \n\t" |
2052 |
|
// "vaddsubpd %%ymm3, %%ymm13, %%ymm13 \n\t" |
2053 |
|
// " \n\t" |
2054 |
|
// " \n\t" |
2055 |
|
// " \n\t" |
2056 |
|
// " \n\t" |
2057 |
|
// " \n\t" |
2058 |
|
// "movq %5, %%rbx \n\t" // load address of beta |
2059 |
|
// "vbroadcastsd (%%rbx), %%ymm1 \n\t" // load beta_r and duplicate |
2060 |
|
// "vbroadcastsd 8(%%rbx), %%ymm2 \n\t" // load beta_i and duplicate |
2061 |
|
// " \n\t" |
2062 |
|
// " \n\t" |
2063 |
|
// " \n\t" |
2064 |
|
// " \n\t" |
2065 |
|
// "movq %7, %%rsi \n\t" // load rs_c |
2066 |
|
// "leaq (,%%rsi,8), %%rsi \n\t" // rsi = rs_c * sizeof(dcomplex) |
2067 |
|
// "leaq (,%%rsi,2), %%rsi \n\t" |
2068 |
|
// "leaq (,%%rsi,2), %%rdx \n\t" // rdx = 2*rs_c; |
2069 |
|
// " \n\t" |
2070 |
|
// " \n\t" |
2071 |
|
// " \n\t" |
2072 |
|
// " \n\t" // now avoid loading C if beta == 0 |
2073 |
|
// "vxorpd %%ymm0, %%ymm0, %%ymm0 \n\t" // set ymm0 to zero. |
2074 |
|
// "vucomisd %%xmm0, %%xmm1 \n\t" // set ZF if beta_r == 0. |
2075 |
|
// "sete %%r8b \n\t" // r8b = ( ZF == 1 ? 1 : 0 ); |
2076 |
|
// "vucomisd %%xmm0, %%xmm2 \n\t" // set ZF if beta_i == 0. |
2077 |
|
// "sete %%r9b \n\t" // r9b = ( ZF == 1 ? 1 : 0 ); |
2078 |
|
// "andb %%r8b, %%r9b \n\t" // set ZF if r8b & r9b == 1. |
2079 |
|
// "jne .ZBETAZERO \n\t" // if ZF = 1, jump to beta == 0 case |
2080 |
|
// " \n\t" |
2081 |
|
// " \n\t" |
2082 |
|
// "cmpq $16, %%rsi \n\t" // set ZF if (16*rs_c) == 16. |
2083 |
|
// "jz .ZCOLSTORED \n\t" // jump to row storage case |
2084 |
|
// " \n\t" |
2085 |
|
// " \n\t" |
2086 |
|
// " \n\t" |
2087 |
|
// ".ZGENSTORED: \n\t" |
2088 |
|
// " \n\t" |
2089 |
|
// " \n\t" |
2090 |
|
// ZGEMM_INPUT_SCALE_GS_BETA_NZ |
2091 |
|
// "vaddpd %%ymm4, %%ymm0, %%ymm0 \n\t" |
2092 |
|
// ZGEMM_OUTPUT_GS |
2093 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2094 |
|
// " \n\t" |
2095 |
|
// " \n\t" |
2096 |
|
// ZGEMM_INPUT_SCALE_GS_BETA_NZ |
2097 |
|
// "vaddpd %%ymm5, %%ymm0, %%ymm0 \n\t" |
2098 |
|
// ZGEMM_OUTPUT_GS |
2099 |
|
// "movq %%r11, %%rcx \n\t" // rcx = c + 1*cs_c |
2100 |
|
// " \n\t" |
2101 |
|
// " \n\t" |
2102 |
|
// " \n\t" |
2103 |
|
// ZGEMM_INPUT_SCALE_GS_BETA_NZ |
2104 |
|
// "vaddpd %%ymm8, %%ymm0, %%ymm0 \n\t" |
2105 |
|
// ZGEMM_OUTPUT_GS |
2106 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2107 |
|
// " \n\t" |
2108 |
|
// " \n\t" |
2109 |
|
// ZGEMM_INPUT_SCALE_GS_BETA_NZ |
2110 |
|
// "vaddpd %%ymm9, %%ymm0, %%ymm0 \n\t" |
2111 |
|
// ZGEMM_OUTPUT_GS |
2112 |
|
// "movq %%r12, %%rcx \n\t" // rcx = c + 2*cs_c |
2113 |
|
// " \n\t" |
2114 |
|
// " \n\t" |
2115 |
|
// " \n\t" |
2116 |
|
// ZGEMM_INPUT_SCALE_GS_BETA_NZ |
2117 |
|
// "vaddpd %%ymm12, %%ymm0, %%ymm0 \n\t" |
2118 |
|
// ZGEMM_OUTPUT_GS |
2119 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2120 |
|
// " \n\t" |
2121 |
|
// " \n\t" |
2122 |
|
// ZGEMM_INPUT_SCALE_GS_BETA_NZ |
2123 |
|
// "vaddpd %%ymm13, %%ymm0, %%ymm0 \n\t" |
2124 |
|
// ZGEMM_OUTPUT_GS |
2125 |
|
// " \n\t" |
2126 |
|
// " \n\t" |
2127 |
|
// " \n\t" |
2128 |
|
// "jmp .ZDONE \n\t" // jump to end. |
2129 |
|
// " \n\t" |
2130 |
|
// " \n\t" |
2131 |
|
// " \n\t" |
2132 |
|
// ".ZCOLSTORED: \n\t" |
2133 |
|
// " \n\t" |
2134 |
|
// " \n\t" |
2135 |
|
// ZGEMM_INPUT_SCALE_CS_BETA_NZ |
2136 |
|
// "vaddpd %%ymm4, %%ymm0, %%ymm0 \n\t" |
2137 |
|
// ZGEMM_OUTPUT_CS |
2138 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2139 |
|
// " \n\t" |
2140 |
|
// " \n\t" |
2141 |
|
// ZGEMM_INPUT_SCALE_CS_BETA_NZ |
2142 |
|
// "vaddpd %%ymm5, %%ymm0, %%ymm0 \n\t" |
2143 |
|
// ZGEMM_OUTPUT_CS |
2144 |
|
// "movq %%r11, %%rcx \n\t" // rcx = c + 1*cs_c |
2145 |
|
// " \n\t" |
2146 |
|
// " \n\t" |
2147 |
|
// " \n\t" |
2148 |
|
// ZGEMM_INPUT_SCALE_CS_BETA_NZ |
2149 |
|
// "vaddpd %%ymm8, %%ymm0, %%ymm0 \n\t" |
2150 |
|
// ZGEMM_OUTPUT_CS |
2151 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2152 |
|
// " \n\t" |
2153 |
|
// " \n\t" |
2154 |
|
// ZGEMM_INPUT_SCALE_CS_BETA_NZ |
2155 |
|
// "vaddpd %%ymm9, %%ymm0, %%ymm0 \n\t" |
2156 |
|
// ZGEMM_OUTPUT_CS |
2157 |
|
// "movq %%r12, %%rcx \n\t" // rcx = c + 2*cs_c |
2158 |
|
// " \n\t" |
2159 |
|
// " \n\t" |
2160 |
|
// " \n\t" |
2161 |
|
// ZGEMM_INPUT_SCALE_CS_BETA_NZ |
2162 |
|
// "vaddpd %%ymm12, %%ymm0, %%ymm0 \n\t" |
2163 |
|
// ZGEMM_OUTPUT_CS |
2164 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2165 |
|
// " \n\t" |
2166 |
|
// " \n\t" |
2167 |
|
// ZGEMM_INPUT_SCALE_CS_BETA_NZ |
2168 |
|
// "vaddpd %%ymm13, %%ymm0, %%ymm0 \n\t" |
2169 |
|
// ZGEMM_OUTPUT_CS |
2170 |
|
// " \n\t" |
2171 |
|
// " \n\t" |
2172 |
|
// " \n\t" |
2173 |
|
// "jmp .ZDONE \n\t" // jump to end. |
2174 |
|
// " \n\t" |
2175 |
|
// " \n\t" |
2176 |
|
// " \n\t" |
2177 |
|
// ".ZBETAZERO: \n\t" |
2178 |
|
// " \n\t" |
2179 |
|
// "cmpq $16, %%rsi \n\t" // set ZF if (16*rs_c) == 16. |
2180 |
|
// "jz .ZCOLSTORBZ \n\t" // jump to row storage case |
2181 |
|
// " \n\t" |
2182 |
|
// " \n\t" |
2183 |
|
// " \n\t" |
2184 |
|
// ".ZGENSTORBZ: \n\t" |
2185 |
|
// " \n\t" |
2186 |
|
// " \n\t" |
2187 |
|
// "vmovaps %%ymm4, %%ymm0 \n\t" |
2188 |
|
// ZGEMM_OUTPUT_GS |
2189 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2190 |
|
// " \n\t" |
2191 |
|
// " \n\t" |
2192 |
|
// "vmovaps %%ymm5, %%ymm0 \n\t" |
2193 |
|
// ZGEMM_OUTPUT_GS |
2194 |
|
// "movq %%r11, %%rcx \n\t" // rcx = c + 1*cs_c |
2195 |
|
// " \n\t" |
2196 |
|
// " \n\t" |
2197 |
|
// " \n\t" |
2198 |
|
// "vmovaps %%ymm8, %%ymm0 \n\t" |
2199 |
|
// ZGEMM_OUTPUT_GS |
2200 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2201 |
|
// " \n\t" |
2202 |
|
// " \n\t" |
2203 |
|
// "vmovaps %%ymm9, %%ymm0 \n\t" |
2204 |
|
// ZGEMM_OUTPUT_GS |
2205 |
|
// "movq %%r12, %%rcx \n\t" // rcx = c + 2*cs_c |
2206 |
|
// " \n\t" |
2207 |
|
// " \n\t" |
2208 |
|
// " \n\t" |
2209 |
|
// "vmovaps %%ymm12, %%ymm0 \n\t" |
2210 |
|
// ZGEMM_OUTPUT_GS |
2211 |
|
// "addq %%rdx, %%rcx \n\t" // c += 2*rs_c; |
2212 |
|
// " \n\t" |
2213 |
|
// " \n\t" |
2214 |
|
// "vmovaps %%ymm13, %%ymm0 \n\t" |
2215 |
|
// ZGEMM_OUTPUT_GS |
2216 |
|
// " \n\t" |
2217 |
|
// " \n\t" |
2218 |
|
// " \n\t" |
2219 |
|
// "jmp .ZDONE \n\t" // jump to end. |
2220 |
|
// " \n\t" |
2221 |
|
// " \n\t" |
2222 |
|
// " \n\t" |
2223 |
|
// ".ZCOLSTORBZ: \n\t" |
2224 |
|
// " \n\t" |
2225 |
|
// " \n\t" |
2226 |
|
// "vmovups %%ymm4, (%%rcx) \n\t" |
2227 |
|
// "vmovups %%ymm5, (%%rcx,%%rdx,1) \n\t" |
2228 |
|
// " \n\t" |
2229 |
|
// "vmovups %%ymm8, (%%r11) \n\t" |
2230 |
|
// "vmovups %%ymm9, (%%r11,%%rdx,1) \n\t" |
2231 |
|
// " \n\t" |
2232 |
|
// "vmovups %%ymm12, (%%r12) \n\t" |
2233 |
|
// "vmovups %%ymm13, (%%r12,%%rdx,1) \n\t" |
2234 |
|
// " \n\t" |
2235 |
|
// " \n\t" |
2236 |
|
// " \n\t" |
2237 |
|
// " \n\t" |
2238 |
|
// " \n\t" |
2239 |
|
// " \n\t" |
2240 |
|
// ".ZDONE: \n\t" |
2241 |
|
// " \n\t" |
2242 |
|
// "vzeroupper \n\t" |
2243 |
|
// " \n\t" |
2244 |
|
// |
2245 |
|
// : // output operands (none) |
2246 |
|
// : // input operands |
2247 |
|
// "m" (k_iter), // 0 |
2248 |
|
// "m" (k_left), // 1 |
2249 |
|
// "m" (a), // 2 |
2250 |
|
// "m" (b), // 3 |
2251 |
|
// "m" (alpha), // 4 |
2252 |
|
// "m" (beta), // 5 |
2253 |
|
// "m" (c), // 6 |
2254 |
|
// "m" (rs_c), // 7 |
2255 |
|
// "m" (cs_c)/*, // 8 |
2256 |
|
// "m" (b_next), // 9 |
2257 |
|
// "m" (a_next)*/ // 10 |
2258 |
|
// : // register clobber list |
2259 |
|
// "rax", "rbx", "rcx", "rdx", "rsi", "rdi", |
2260 |
|
// "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15", |
2261 |
|
// "xmm0", "xmm1", "xmm2", "xmm3", |
2262 |
|
// "xmm4", "xmm5", "xmm6", "xmm7", |
2263 |
|
// "xmm8", "xmm9", "xmm10", "xmm11", |
2264 |
|
// "xmm12", "xmm13", "xmm14", "xmm15", |
2265 |
|
// "memory" |
2266 |
|
// ); |
2267 |
|
//} |
2268 |
|
// |