1 /***************************************************************************
2 Copyright (c) 2014, 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
29 /* need a new enough GCC for avx512 support */
30 #if (( defined(__GNUC__) && __GNUC__ > 6 && defined(__AVX2__)) || (defined(__clang__) && __clang_major__ >= 6))
31
32 #include <immintrin.h>
33
34 #define HAVE_KERNEL_4x4 1
35
36 #if defined(__clang_patchlevel__) && __clang_major__ == 9 && __clang_minor__ == 0 && __clang_patchlevel__ == 0
37 #pragma clang optimize off
38 #endif
39 #if defined(__apple_build_version__) && __clang_major__ == 11 && __clang_minor__ == 0 && __clang_patchlevel__ == 3
40 #pragma clang optimize off
41 #endif
dsymv_kernel_4x4(BLASLONG from,BLASLONG to,FLOAT ** a,FLOAT * x,FLOAT * y,FLOAT * temp1,FLOAT * temp2)42 static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FLOAT *y, FLOAT *temp1, FLOAT *temp2)
43 {
44
45
46 __m256d accum_0, accum_1, accum_2, accum_3;
47 __m256d temp1_0, temp1_1, temp1_2, temp1_3;
48
49 /* the 256 bit wide acculmulator vectors start out as zero */
50 accum_0 = _mm256_setzero_pd();
51 accum_1 = _mm256_setzero_pd();
52 accum_2 = _mm256_setzero_pd();
53 accum_3 = _mm256_setzero_pd();
54
55 temp1_0 = _mm256_broadcastsd_pd(_mm_load_sd(&temp1[0]));
56 temp1_1 = _mm256_broadcastsd_pd(_mm_load_sd(&temp1[1]));
57 temp1_2 = _mm256_broadcastsd_pd(_mm_load_sd(&temp1[2]));
58 temp1_3 = _mm256_broadcastsd_pd(_mm_load_sd(&temp1[3]));
59
60 #ifdef __AVX512CD__
61 __m512d accum_05, accum_15, accum_25, accum_35;
62 __m512d temp1_05, temp1_15, temp1_25, temp1_35;
63 BLASLONG to2;
64 int delta;
65
66 /* the 512 bit wide accumulator vectors start out as zero */
67 accum_05 = _mm512_setzero_pd();
68 accum_15 = _mm512_setzero_pd();
69 accum_25 = _mm512_setzero_pd();
70 accum_35 = _mm512_setzero_pd();
71
72 temp1_05 = _mm512_broadcastsd_pd(_mm_load_sd(&temp1[0]));
73 temp1_15 = _mm512_broadcastsd_pd(_mm_load_sd(&temp1[1]));
74 temp1_25 = _mm512_broadcastsd_pd(_mm_load_sd(&temp1[2]));
75 temp1_35 = _mm512_broadcastsd_pd(_mm_load_sd(&temp1[3]));
76
77 delta = (to - from) & ~7;
78 to2 = from + delta;
79
80
81 for (; from < to2; from += 8) {
82 __m512d _x, _y;
83 __m512d a0, a1, a2, a3;
84
85 _y = _mm512_loadu_pd(&y[from]);
86 _x = _mm512_loadu_pd(&x[from]);
87
88 a0 = _mm512_loadu_pd(&a[0][from]);
89 a1 = _mm512_loadu_pd(&a[1][from]);
90 a2 = _mm512_loadu_pd(&a[2][from]);
91 a3 = _mm512_loadu_pd(&a[3][from]);
92
93 _y += temp1_05 * a0 + temp1_15 * a1 + temp1_25 * a2 + temp1_35 * a3;
94
95 accum_05 += _x * a0;
96 accum_15 += _x * a1;
97 accum_25 += _x * a2;
98 accum_35 += _x * a3;
99
100 _mm512_storeu_pd(&y[from], _y);
101
102 };
103
104 /*
105 * we need to fold our 512 bit wide accumulator vectors into 256 bit wide vectors so that the AVX2 code
106 * below can continue using the intermediate results in its loop
107 */
108 accum_0 = _mm256_add_pd(_mm512_extractf64x4_pd(accum_05, 0), _mm512_extractf64x4_pd(accum_05, 1));
109 accum_1 = _mm256_add_pd(_mm512_extractf64x4_pd(accum_15, 0), _mm512_extractf64x4_pd(accum_15, 1));
110 accum_2 = _mm256_add_pd(_mm512_extractf64x4_pd(accum_25, 0), _mm512_extractf64x4_pd(accum_25, 1));
111 accum_3 = _mm256_add_pd(_mm512_extractf64x4_pd(accum_35, 0), _mm512_extractf64x4_pd(accum_35, 1));
112
113 #endif
114
115 for (; from != to; from += 4) {
116 __m256d _x, _y;
117 __m256d a0, a1, a2, a3;
118
119 _y = _mm256_loadu_pd(&y[from]);
120 _x = _mm256_loadu_pd(&x[from]);
121
122 /* load 4 rows of matrix data */
123 a0 = _mm256_loadu_pd(&a[0][from]);
124 a1 = _mm256_loadu_pd(&a[1][from]);
125 a2 = _mm256_loadu_pd(&a[2][from]);
126 a3 = _mm256_loadu_pd(&a[3][from]);
127
128 _y += temp1_0 * a0 + temp1_1 * a1 + temp1_2 * a2 + temp1_3 * a3;
129
130 accum_0 += _x * a0;
131 accum_1 += _x * a1;
132 accum_2 += _x * a2;
133 accum_3 += _x * a3;
134
135 _mm256_storeu_pd(&y[from], _y);
136
137 };
138
139 /*
140 * we now have 4 accumulator vectors. Each vector needs to be summed up element wise and stored in the temp2
141 * output array. There is no direct instruction for this in 256 bit space, only in 128 space.
142 */
143
144 __m128d half_accum0, half_accum1, half_accum2, half_accum3;
145
146
147 /* Add upper half to lower half of each of the four 256 bit vectors to get to four 128 bit vectors */
148 half_accum0 = _mm_add_pd(_mm256_extractf128_pd(accum_0, 0), _mm256_extractf128_pd(accum_0, 1));
149 half_accum1 = _mm_add_pd(_mm256_extractf128_pd(accum_1, 0), _mm256_extractf128_pd(accum_1, 1));
150 half_accum2 = _mm_add_pd(_mm256_extractf128_pd(accum_2, 0), _mm256_extractf128_pd(accum_2, 1));
151 half_accum3 = _mm_add_pd(_mm256_extractf128_pd(accum_3, 0), _mm256_extractf128_pd(accum_3, 1));
152
153 /* in 128 bit land there is a hadd operation to do the rest of the element-wise sum in one go */
154 half_accum0 = _mm_hadd_pd(half_accum0, half_accum0);
155 half_accum1 = _mm_hadd_pd(half_accum1, half_accum1);
156 half_accum2 = _mm_hadd_pd(half_accum2, half_accum2);
157 half_accum3 = _mm_hadd_pd(half_accum3, half_accum3);
158
159 /* and store the lowest double value from each of these vectors in the temp2 output */
160 temp2[0] += half_accum0[0];
161 temp2[1] += half_accum1[0];
162 temp2[2] += half_accum2[0];
163 temp2[3] += half_accum3[0];
164 }
165
166 #if defined(__clang_patchlevel__) && __clang_major__ == 9 && __clang_minor__ == 0 && __clang_patchlevel__ == 0
167 #pragma clang optimize on
168 #endif
169 #if defined(__apple_build_version__) && __clang_major__ == 11 && __clang_minor__ == 0 && __clang_patchlevel__ == 3
170 #pragma clang optimize on
171 #endif
172
173 #else
174 #include "dsymv_L_microk_haswell-2.c"
175 #endif
176