1 /***************************************************************************
2 Copyright (c) 2017, The OpenBLAS Project
3 All rights reserved.
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions are
6 met:
7 1. Redistributions of source code must retain the above copyright
8 notice, this list of conditions and the following disclaimer.
9 2. Redistributions in binary form must reproduce the above copyright
10 notice, this list of conditions and the following disclaimer in
11 the documentation and/or other materials provided with the
12 distribution.
13 3. Neither the name of the OpenBLAS project nor the names of
14 its contributors may be used to endorse or promote products
15 derived from this software without specific prior written permission.
16 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
20 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24 OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
25 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 *****************************************************************************/
27
28 #include "common.h"
29
30 #include <arm_neon.h>
31
32 #define N "x0" /* vector length */
33 #define X "x1" /* "X" vector address */
34 #define INC_X "x2" /* "X" stride */
35 #define J "x5" /* loop variable */
36
37 #define REG0 "wzr"
38 #define SUMF "s0"
39 #define SUMFD "d0"
40
41 /******************************************************************************/
42
43 #define KERNEL_F1 \
44 "ldr d1, ["X"] \n" \
45 "add "X", "X", #8 \n" \
46 "fabs v1.2s, v1.2s \n" \
47 "ext v2.8b, v1.8b, v1.8b, #4 \n" \
48 "fadd s1, s1, s2 \n" \
49 "fadd "SUMF", "SUMF", s1 \n"
50
51 #define KERNEL_F32 \
52 "ldr q16, ["X"] \n" \
53 "ldr q17, ["X", #16] \n" \
54 "ldr q18, ["X", #32] \n" \
55 "ldr q19, ["X", #48] \n" \
56 "ldp q20, q21, ["X", #64] \n" \
57 "ldp q22, q23, ["X", #96] \n" \
58 "fabs v16.4s, v16.4s \n" \
59 "fabs v17.4s, v17.4s \n" \
60 "fabs v18.4s, v18.4s \n" \
61 "fabs v19.4s, v19.4s \n" \
62 "ldp q24, q25, ["X", #128] \n" \
63 "ldp q26, q27, ["X", #160] \n" \
64 "fabs v20.4s, v20.4s \n" \
65 "fabs v21.4s, v21.4s \n" \
66 "fabs v22.4s, v22.4s \n" \
67 "fabs v23.4s, v23.4s \n" \
68 "fadd v16.4s, v16.4s, v17.4s \n" \
69 "fadd v18.4s, v18.4s, v19.4s \n" \
70 "ldp q28, q29, ["X", #192] \n" \
71 "ldp q30, q31, ["X", #224] \n" \
72 "fabs v24.4s, v24.4s \n" \
73 "fabs v25.4s, v25.4s \n" \
74 "fabs v26.4s, v26.4s \n" \
75 "fabs v27.4s, v27.4s \n" \
76 "add "X", "X", #256 \n" \
77 "fadd v20.4s, v20.4s, v21.4s \n" \
78 "fadd v22.4s, v22.4s, v23.4s \n" \
79 "fabs v28.4s, v28.4s \n" \
80 "fabs v29.4s, v29.4s \n" \
81 "fabs v30.4s, v30.4s \n" \
82 "fabs v31.4s, v31.4s \n" \
83 "PRFM PLDL1KEEP, ["X", #1024] \n" \
84 "PRFM PLDL1KEEP, ["X", #1024+64] \n" \
85 "fadd v24.4s, v24.4s, v25.4s \n" \
86 "fadd v26.4s, v26.4s, v27.4s \n" \
87 "fadd v0.4s, v0.4s, v16.4s \n" \
88 "fadd v1.4s, v1.4s, v18.4s \n" \
89 "fadd v2.4s, v2.4s, v20.4s \n" \
90 "fadd v3.4s, v3.4s, v22.4s \n" \
91 "PRFM PLDL1KEEP, ["X", #1024+128] \n" \
92 "PRFM PLDL1KEEP, ["X", #1024+192] \n" \
93 "fadd v28.4s, v28.4s, v29.4s \n" \
94 "fadd v30.4s, v30.4s, v31.4s \n" \
95 "fadd v4.4s, v4.4s, v24.4s \n" \
96 "fadd v5.4s, v5.4s, v26.4s \n" \
97 "fadd v6.4s, v6.4s, v28.4s \n" \
98 "fadd v7.4s, v7.4s, v30.4s \n"
99
100 #define KERNEL_F32_FINALIZE \
101 "fadd v0.4s, v0.4s, v1.4s \n" \
102 "fadd v2.4s, v2.4s, v3.4s \n" \
103 "fadd v4.4s, v4.4s, v5.4s \n" \
104 "fadd v6.4s, v6.4s, v7.4s \n" \
105 "fadd v0.4s, v0.4s, v2.4s \n" \
106 "fadd v4.4s, v4.4s, v6.4s \n" \
107 "fadd v0.4s, v0.4s, v4.4s \n" \
108 "ext v1.16b, v0.16b, v0.16b, #8 \n" \
109 "fadd v0.2s, v0.2s, v1.2s \n" \
110 "faddp "SUMF", v0.2s \n"
111
112 #define INIT_S \
113 "lsl "INC_X", "INC_X", #3 \n"
114
115 #define KERNEL_S1 \
116 "ldr d1, ["X"] \n" \
117 "add "X", "X", "INC_X" \n" \
118 "fabs v1.2s, v1.2s \n" \
119 "ext v2.8b, v1.8b, v1.8b, #4 \n" \
120 "fadd s1, s1, s2 \n" \
121 "fadd "SUMF", "SUMF", s1 \n"
122
123
124 #if defined(SMP)
125 extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n,
126 BLASLONG k, void *alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb,
127 void *c, BLASLONG ldc, int (*function)(), int nthreads);
128 #endif
129
130
casum_compute(BLASLONG n,FLOAT * x,BLASLONG inc_x)131 static FLOAT casum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
132 {
133 FLOAT asum = 0.0 ;
134
135 if ( n < 0 ) return(asum);
136
137 __asm__ __volatile__ (
138 " mov "N", %[N_] \n"
139 " mov "X", %[X_] \n"
140 " mov "INC_X", %[INCX_] \n"
141 " fmov "SUMF", "REG0" \n"
142 " fmov s1, "REG0" \n"
143 " fmov s2, "REG0" \n"
144 " fmov s3, "REG0" \n"
145 " fmov s4, "REG0" \n"
146 " fmov s5, "REG0" \n"
147 " fmov s6, "REG0" \n"
148 " fmov s7, "REG0" \n"
149 " cmp "N", xzr \n"
150 " ble 9f //asum_kernel_L999 \n"
151 " cmp "INC_X", xzr \n"
152 " ble 9f //asum_kernel_L999 \n"
153 " cmp "INC_X", #1 \n"
154 " bne 5f //asum_kernel_S_BEGIN \n"
155
156 "1: //asum_kernel_F_BEGIN: \n"
157 " asr "J", "N", #5 \n"
158 " cmp "J", xzr \n"
159 " beq 3f //asum_kernel_F1 \n"
160
161 "2: //asum_kernel_F32: \n"
162 " "KERNEL_F32" \n"
163 " subs "J", "J", #1 \n"
164 " bne 2b //asum_kernel_F32 \n"
165 " "KERNEL_F32_FINALIZE" \n"
166
167 "3: //asum_kernel_F1: \n"
168 " ands "J", "N", #31 \n"
169 " ble 9f //asum_kernel_L999 \n"
170
171 "4: //asum_kernel_F10: \n"
172 " "KERNEL_F1" \n"
173 " subs "J", "J", #1 \n"
174 " bne 4b //asum_kernel_F10 \n"
175 " b 9f //asum_kernel_L999 \n"
176
177 "5: //asum_kernel_S_BEGIN: \n"
178 " "INIT_S" \n"
179 " asr "J", "N", #2 \n"
180 " cmp "J", xzr \n"
181 " ble 7f //asum_kernel_S1 \n"
182
183 "6: //asum_kernel_S4: \n"
184 " "KERNEL_S1" \n"
185 " "KERNEL_S1" \n"
186 " "KERNEL_S1" \n"
187 " "KERNEL_S1" \n"
188 " subs "J", "J", #1 \n"
189 " bne 6b //asum_kernel_S4 \n"
190
191 "7: //asum_kernel_S1: \n"
192 " ands "J", "N", #3 \n"
193 " ble 9f //asum_kernel_L999 \n"
194
195 "8: //asum_kernel_S10: \n"
196 " "KERNEL_S1" \n"
197 " subs "J", "J", #1 \n"
198 " bne 8b //asum_kernel_S10 \n"
199
200 "9: //asum_kernel_L999: \n"
201 " fmov %[ASUM_], "SUMFD" \n"
202
203 : [ASUM_] "=r" (asum) //%0
204 : [N_] "r" (n), //%1
205 [X_] "r" (x), //%2
206 [INCX_] "r" (inc_x) //%3
207 : "cc",
208 "memory",
209 "x0", "x1", "x2", "x3", "x4", "x5",
210 "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"
211 );
212
213 return asum;
214 }
215
216 #if defined(SMP)
casum_thread_function(BLASLONG n,BLASLONG dummy0,BLASLONG dummy1,FLOAT dummy2,FLOAT * x,BLASLONG inc_x,FLOAT * y,BLASLONG inc_y,FLOAT * result,BLASLONG dummy3)217 static int casum_thread_function(BLASLONG n, BLASLONG dummy0,
218 BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
219 BLASLONG inc_y, FLOAT *result, BLASLONG dummy3)
220 {
221 *result = casum_compute(n, x, inc_x);
222
223 return 0;
224 }
225 #endif
226
CNAME(BLASLONG n,FLOAT * x,BLASLONG inc_x)227 FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
228 {
229 #if defined(SMP)
230 int nthreads;
231 FLOAT dummy_alpha;
232 #endif
233 FLOAT asum = 0.0;
234
235 #if defined(SMP)
236 if (inc_x == 0 || n <= 10000)
237 nthreads = 1;
238 else
239 nthreads = num_cpu_avail(1);
240
241 if (nthreads == 1) {
242 asum = casum_compute(n, x, inc_x);
243 } else {
244 int mode, i;
245 char result[MAX_CPU_NUMBER * sizeof(double) * 2];
246 FLOAT *ptr;
247
248 mode = BLAS_SINGLE | BLAS_COMPLEX;
249
250 blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
251 x, inc_x, NULL, 0, result, 0,
252 ( void *)casum_thread_function, nthreads);
253
254 ptr = (FLOAT *)result;
255 for (i = 0; i < nthreads; i++) {
256 asum = asum + (*ptr);
257 ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
258 }
259 }
260 #else
261 asum = casum_compute(n, x, inc_x);
262 #endif
263
264 return asum;
265 }
266