hw1: small allocation fix on ex2

This commit is contained in:
Claudio Maggioni 2022-10-03 21:44:34 +02:00
parent ce33ca6d7a
commit 8013c535e6
5 changed files with 231 additions and 231 deletions

View File

@ -17,7 +17,7 @@ LDLIBS = -lrt -Wl,--start-group $(MKLROOT)/lib/intel64/libmkl_intel_lp64.a $(MKL
const char* dgemm_desc = "Block-based dgemm.";
const int block_size = 27;
const int block_size = 26;
inline int min(int a, int b) {
return a < b ? a : b;
@ -45,10 +45,10 @@ inline void store_c(double* C, double* C_temp, int r_min, int r_max, int c_min,
* On exit, A and B maintain their input values. */
void square_dgemm(int n, double* A, double* B, double* C) {
double A_row[n * n];
double row_tmp[n];
double C_temp[block_size * block_size];
for (int m = 0; m < n; ++m) {
double row_tmp[n];
memcpy(row_tmp, A + m * n, n * sizeof(double));
for (int l = 0; l < n; ++l) {

View File

@ -1,7 +1,7 @@
%!PS-Adobe-2.0
%%Title: timing.ps
%%Creator: gnuplot 5.2 patchlevel 8
%%CreationDate: Mon Oct 3 21:19:18 2022
%%CreationDate: Mon Oct 3 21:44:11 2022
%%DocumentFonts: (atend)
%%BoundingBox: 50 50 554 770
%%Orientation: Landscape
@ -483,7 +483,7 @@ SDict begin [
/Creator (gnuplot 5.2 patchlevel 8)
% /Producer (gnuplot)
% /Keywords ()
/CreationDate (Mon Oct 3 21:19:18 2022)
/CreationDate (Mon Oct 3 21:44:11 2022)
/DOCINFO pdfmark
end
} ifelse
@ -961,57 +961,57 @@ LTb
6380 4486 M
399 0 V
1171 2365 M
7 -1 V
481 -108 V
8 -1 V
225 -22 V
7 -72 V
8 71 V
466 -51 V
7 -29 V
278 23 V
195 0 V
8 -472 V
7 469 V
466 -6 V
8 -69 V
7 77 V
721 -2 V
466 -5 V
7 -46 V
233 45 V
8 -473 V
954 464 V
7 -451 V
954 450 V
8 -550 V
7 546 V
7 2 V
481 -111 V
8 0 V
225 -23 V
7 -63 V
8 57 V
466 -55 V
7 -15 V
278 7 V
195 2 V
8 -480 V
7 483 V
466 -9 V
8 -112 V
7 125 V
721 -3 V
466 0 V
7 -19 V
233 20 V
8 -480 V
954 473 V
7 -573 V
954 564 V
8 -457 V
7 460 V
1171 2365 Pls
1178 2364 Pls
1178 2367 Pls
1659 2256 Pls
1667 2255 Pls
1667 2256 Pls
1892 2233 Pls
1899 2161 Pls
1907 2232 Pls
2373 2181 Pls
2380 2152 Pls
2658 2175 Pls
2853 2175 Pls
2861 1703 Pls
2868 2172 Pls
3334 2166 Pls
3342 2097 Pls
3349 2174 Pls
4070 2172 Pls
4536 2167 Pls
4543 2121 Pls
4776 2166 Pls
4784 1693 Pls
5738 2157 Pls
5745 1706 Pls
6699 2156 Pls
6707 1606 Pls
6714 2152 Pls
1899 2170 Pls
1907 2227 Pls
2373 2172 Pls
2380 2157 Pls
2658 2164 Pls
2853 2166 Pls
2861 1686 Pls
2868 2169 Pls
3334 2160 Pls
3342 2048 Pls
3349 2173 Pls
4070 2170 Pls
4536 2170 Pls
4543 2151 Pls
4776 2171 Pls
4784 1691 Pls
5738 2164 Pls
5745 1591 Pls
6699 2155 Pls
6707 1698 Pls
6714 2158 Pls
6579 4486 Pls
% End plot #1
% Begin plot #2
@ -1030,58 +1030,58 @@ LTb
0.00 0.62 0.45 C
6380 4346 M
399 0 V
1171 2307 M
7 13 V
481 34 V
8 4 V
225 4 V
7 -53 V
8 21 V
466 17 V
7 4 V
278 -3 V
195 16 V
8 -17 V
7 1 V
466 -7 V
8 23 V
7 5 V
721 -1 V
466 -24 V
7 13 V
233 9 V
8 2 V
954 -1 V
7 2 V
954 -4 V
8 -9 V
7 4 V
1171 2307 Crs
1178 2320 Crs
1659 2354 Crs
1667 2358 Crs
1892 2362 Crs
1899 2309 Crs
1907 2330 Crs
2373 2347 Crs
2380 2351 Crs
2658 2348 Crs
2853 2364 Crs
2861 2347 Crs
2868 2348 Crs
3334 2341 Crs
3342 2364 Crs
3349 2369 Crs
4070 2368 Crs
4536 2344 Crs
4543 2357 Crs
4776 2366 Crs
4784 2368 Crs
5738 2367 Crs
5745 2369 Crs
6699 2365 Crs
6707 2356 Crs
6714 2360 Crs
1171 2290 M
7 21 V
481 37 V
8 -1 V
225 2 V
7 -12 V
8 13 V
466 3 V
7 0 V
278 2 V
195 -5 V
8 -7 V
7 6 V
466 8 V
8 1 V
7 0 V
721 3 V
466 -38 V
7 -12 V
233 10 V
8 16 V
954 2 V
7 -7 V
954 16 V
8 10 V
7 -11 V
1171 2290 Crs
1178 2311 Crs
1659 2348 Crs
1667 2347 Crs
1892 2349 Crs
1899 2337 Crs
1907 2350 Crs
2373 2353 Crs
2380 2353 Crs
2658 2355 Crs
2853 2350 Crs
2861 2343 Crs
2868 2349 Crs
3334 2357 Crs
3342 2358 Crs
3349 2358 Crs
4070 2361 Crs
4536 2323 Crs
4543 2311 Crs
4776 2321 Crs
4784 2337 Crs
5738 2339 Crs
5745 2332 Crs
6699 2348 Crs
6707 2358 Crs
6714 2347 Crs
6579 4346 Crs
% End plot #2
% Begin plot #3
@ -1100,58 +1100,58 @@ LTb
0.34 0.71 0.91 C
6380 4206 M
399 0 V
1171 3798 M
7 71 V
481 76 V
8 -23 V
225 -32 V
7 49 V
8 -22 V
466 14 V
7 57 V
278 -14 V
195 -12 V
8 21 V
7 -22 V
1171 3743 M
7 112 V
481 86 V
8 -53 V
225 -26 V
7 81 V
8 -27 V
466 20 V
7 10 V
278 28 V
195 -16 V
8 0 V
7 4 V
466 12 V
8 35 V
7 -16 V
721 8 V
466 4 V
7 29 V
233 -47 V
8 30 V
954 1 V
7 -5 V
954 -15 V
721 11 V
466 -55 V
7 74 V
233 -53 V
8 41 V
7 -22 V
1171 3798 Star
1178 3869 Star
1659 3945 Star
1667 3922 Star
1892 3890 Star
1899 3939 Star
1907 3917 Star
2373 3931 Star
2380 3988 Star
954 -22 V
7 22 V
954 -14 V
8 -64 V
7 17 V
1171 3743 Star
1178 3855 Star
1659 3941 Star
1667 3888 Star
1892 3862 Star
1899 3943 Star
1907 3916 Star
2373 3936 Star
2380 3946 Star
2658 3974 Star
2853 3962 Star
2861 3983 Star
2868 3961 Star
3334 3973 Star
3342 4008 Star
3349 3992 Star
4070 4000 Star
4536 4004 Star
4543 4033 Star
4776 3986 Star
4784 4016 Star
5738 4017 Star
5745 4012 Star
2853 3958 Star
2861 3958 Star
2868 3962 Star
3334 3974 Star
3342 4009 Star
3349 3993 Star
4070 4004 Star
4536 3949 Star
4543 4023 Star
4776 3970 Star
4784 4011 Star
5738 3989 Star
5745 4011 Star
6699 3997 Star
6707 4038 Star
6714 4016 Star
6707 3933 Star
6714 3950 Star
6579 4206 Star
% End plot #3
2.000 UL

View File

@ -1,29 +1,29 @@
#Description: Naive, three-loop dgemm.
Size: 31 Mflop/s: 2392.32 Percentage: 6.50
Size: 32 Mflop/s: 2389.31 Percentage: 6.49
Size: 96 Mflop/s: 1996.82 Percentage: 5.43
Size: 97 Mflop/s: 1995.22 Percentage: 5.42
Size: 127 Mflop/s: 1923.52 Percentage: 5.23
Size: 128 Mflop/s: 1707.73 Percentage: 4.64
Size: 129 Mflop/s: 1917.74 Percentage: 5.21
Size: 191 Mflop/s: 1763.08 Percentage: 4.79
Size: 192 Mflop/s: 1681.61 Percentage: 4.57
Size: 229 Mflop/s: 1746.72 Percentage: 4.75
Size: 255 Mflop/s: 1745.99 Percentage: 4.74
Size: 256 Mflop/s: 799.816 Percentage: 2.17
Size: 257 Mflop/s: 1737 Percentage: 4.72
Size: 319 Mflop/s: 1721.21 Percentage: 4.68
Size: 320 Mflop/s: 1535.18 Percentage: 4.17
Size: 321 Mflop/s: 1743.01 Percentage: 4.74
Size: 417 Mflop/s: 1737.5 Percentage: 4.72
Size: 479 Mflop/s: 1724.83 Percentage: 4.69
Size: 480 Mflop/s: 1597.39 Percentage: 4.34
Size: 511 Mflop/s: 1720.23 Percentage: 4.67
Size: 512 Mflop/s: 785.527 Percentage: 2.13
Size: 639 Mflop/s: 1694.76 Percentage: 4.61
Size: 640 Mflop/s: 803.077 Percentage: 2.18
Size: 767 Mflop/s: 1692.02 Percentage: 4.60
Size: 768 Mflop/s: 680.299 Percentage: 1.85
Size: 769 Mflop/s: 1679.99 Percentage: 4.57
#Average percentage of Peak = 4.48494
Size: 31 Mflop/s: 2393.33 Percentage: 6.50
Size: 32 Mflop/s: 2400.13 Percentage: 6.52
Size: 96 Mflop/s: 1998.74 Percentage: 5.43
Size: 97 Mflop/s: 1996.01 Percentage: 5.42
Size: 127 Mflop/s: 1923.81 Percentage: 5.23
Size: 128 Mflop/s: 1731.98 Percentage: 4.71
Size: 129 Mflop/s: 1903.31 Percentage: 5.17
Size: 191 Mflop/s: 1736.78 Percentage: 4.72
Size: 192 Mflop/s: 1694.44 Percentage: 4.60
Size: 229 Mflop/s: 1715.1 Percentage: 4.66
Size: 255 Mflop/s: 1720.39 Percentage: 4.67
Size: 256 Mflop/s: 777.65 Percentage: 2.11
Size: 257 Mflop/s: 1729.27 Percentage: 4.70
Size: 319 Mflop/s: 1704.8 Percentage: 4.63
Size: 320 Mflop/s: 1414.84 Percentage: 3.84
Size: 321 Mflop/s: 1741.3 Percentage: 4.73
Size: 417 Mflop/s: 1733 Percentage: 4.71
Size: 479 Mflop/s: 1731.17 Percentage: 4.70
Size: 480 Mflop/s: 1678.77 Percentage: 4.56
Size: 511 Mflop/s: 1733.6 Percentage: 4.71
Size: 512 Mflop/s: 782.96 Percentage: 2.13
Size: 639 Mflop/s: 1714.42 Percentage: 4.66
Size: 640 Mflop/s: 663.418 Percentage: 1.80
Size: 767 Mflop/s: 1690.82 Percentage: 4.59
Size: 768 Mflop/s: 792.043 Percentage: 2.15
Size: 769 Mflop/s: 1696.95 Percentage: 4.61
#Average percentage of Peak = 4.47314

View File

@ -1,29 +1,29 @@
#Description: Reference dgemm.
Size: 31 Mflop/s: 25682 Percentage: 69.79
Size: 32 Mflop/s: 28883 Percentage: 78.49
Size: 96 Mflop/s: 32725.3 Percentage: 88.93
Size: 97 Mflop/s: 31538.2 Percentage: 85.70
Size: 127 Mflop/s: 29895.3 Percentage: 81.24
Size: 128 Mflop/s: 32447.3 Percentage: 88.17
Size: 129 Mflop/s: 31273.9 Percentage: 84.98
Size: 191 Mflop/s: 32026.4 Percentage: 87.03
Size: 192 Mflop/s: 35140.4 Percentage: 95.49
Size: 229 Mflop/s: 34375.4 Percentage: 93.41
Size: 255 Mflop/s: 33708.5 Percentage: 91.60
Size: 256 Mflop/s: 34891.4 Percentage: 94.81
Size: 257 Mflop/s: 33617.2 Percentage: 91.35
Size: 319 Mflop/s: 34283 Percentage: 93.16
Size: 320 Mflop/s: 36365.1 Percentage: 98.82
Size: 321 Mflop/s: 35399.5 Percentage: 96.19
Size: 417 Mflop/s: 35893.2 Percentage: 97.54
Size: 479 Mflop/s: 36084.1 Percentage: 98.05
Size: 480 Mflop/s: 37868.8 Percentage:102.90
Size: 511 Mflop/s: 35047.8 Percentage: 95.24
Size: 512 Mflop/s: 36868 Percentage:100.18
Size: 639 Mflop/s: 36890.8 Percentage:100.25
Size: 640 Mflop/s: 36603 Percentage: 99.46
Size: 767 Mflop/s: 35669.9 Percentage: 96.93
Size: 768 Mflop/s: 38223.9 Percentage:103.87
Size: 769 Mflop/s: 36856.3 Percentage:100.15
#Average percentage of Peak = 92.8363
Size: 31 Mflop/s: 23449.2 Percentage: 63.72
Size: 32 Mflop/s: 28198.9 Percentage: 76.63
Size: 96 Mflop/s: 32542.3 Percentage: 88.43
Size: 97 Mflop/s: 29801.3 Percentage: 80.98
Size: 127 Mflop/s: 28557.8 Percentage: 77.60
Size: 128 Mflop/s: 32643.3 Percentage: 88.70
Size: 129 Mflop/s: 31198.2 Percentage: 84.78
Size: 191 Mflop/s: 32247.3 Percentage: 87.63
Size: 192 Mflop/s: 32830.6 Percentage: 89.21
Size: 229 Mflop/s: 34360.9 Percentage: 93.37
Size: 255 Mflop/s: 33477.7 Percentage: 90.97
Size: 256 Mflop/s: 33473.9 Percentage: 90.96
Size: 257 Mflop/s: 33686.5 Percentage: 91.54
Size: 319 Mflop/s: 34335.2 Percentage: 93.30
Size: 320 Mflop/s: 36438.1 Percentage: 99.02
Size: 321 Mflop/s: 35433.7 Percentage: 96.29
Size: 417 Mflop/s: 36133.7 Percentage: 98.19
Size: 479 Mflop/s: 32951.4 Percentage: 89.54
Size: 480 Mflop/s: 37260 Percentage:101.25
Size: 511 Mflop/s: 34128 Percentage: 92.74
Size: 512 Mflop/s: 36526.4 Percentage: 99.26
Size: 639 Mflop/s: 35249.2 Percentage: 95.79
Size: 640 Mflop/s: 36538.7 Percentage: 99.29
Size: 767 Mflop/s: 35718.5 Percentage: 97.06
Size: 768 Mflop/s: 32116.8 Percentage: 87.27
Size: 769 Mflop/s: 33033.9 Percentage: 89.77
#Average percentage of Peak = 90.1266

View File

@ -1,29 +1,29 @@
#Description: Block-based dgemm.
Size: 31 Mflop/s: 2174.53 Percentage: 5.91
Size: 32 Mflop/s: 2219.23 Percentage: 6.03
Size: 96 Mflop/s: 2347.86 Percentage: 6.38
Size: 97 Mflop/s: 2366.05 Percentage: 6.43
Size: 127 Mflop/s: 2380.58 Percentage: 6.47
Size: 128 Mflop/s: 2180.15 Percentage: 5.92
Size: 129 Mflop/s: 2257.65 Percentage: 6.13
Size: 191 Mflop/s: 2323.3 Percentage: 6.31
Size: 192 Mflop/s: 2338.71 Percentage: 6.36
Size: 229 Mflop/s: 2327.57 Percentage: 6.32
Size: 255 Mflop/s: 2389.61 Percentage: 6.49
Size: 256 Mflop/s: 2321.28 Percentage: 6.31
Size: 257 Mflop/s: 2327.47 Percentage: 6.32
Size: 319 Mflop/s: 2297.83 Percentage: 6.24
Size: 320 Mflop/s: 2386.76 Percentage: 6.49
Size: 321 Mflop/s: 2408.12 Percentage: 6.54
Size: 417 Mflop/s: 2405.69 Percentage: 6.54
Size: 479 Mflop/s: 2311.13 Percentage: 6.28
Size: 480 Mflop/s: 2359.66 Percentage: 6.41
Size: 511 Mflop/s: 2396.08 Percentage: 6.51
Size: 512 Mflop/s: 2403.92 Percentage: 6.53
Size: 639 Mflop/s: 2401.71 Percentage: 6.53
Size: 640 Mflop/s: 2408.37 Percentage: 6.54
Size: 767 Mflop/s: 2394.15 Percentage: 6.51
Size: 768 Mflop/s: 2358.31 Percentage: 6.41
Size: 769 Mflop/s: 2372.9 Percentage: 6.45
#Average percentage of Peak = 6.36064
Size: 31 Mflop/s: 2112.63 Percentage: 5.74
Size: 32 Mflop/s: 2187.44 Percentage: 5.94
Size: 96 Mflop/s: 2325.39 Percentage: 6.32
Size: 97 Mflop/s: 2322.81 Percentage: 6.31
Size: 127 Mflop/s: 2330.3 Percentage: 6.33
Size: 128 Mflop/s: 2282.93 Percentage: 6.20
Size: 129 Mflop/s: 2334.25 Percentage: 6.34
Size: 191 Mflop/s: 2345.91 Percentage: 6.37
Size: 192 Mflop/s: 2345.38 Percentage: 6.37
Size: 229 Mflop/s: 2351.01 Percentage: 6.39
Size: 255 Mflop/s: 2335.21 Percentage: 6.35
Size: 256 Mflop/s: 2306.48 Percentage: 6.27
Size: 257 Mflop/s: 2330.68 Percentage: 6.33
Size: 319 Mflop/s: 2360.03 Percentage: 6.41
Size: 320 Mflop/s: 2364.53 Percentage: 6.43
Size: 321 Mflop/s: 2366.38 Percentage: 6.43
Size: 417 Mflop/s: 2378.34 Percentage: 6.46
Size: 479 Mflop/s: 2233.05 Percentage: 6.07
Size: 480 Mflop/s: 2187.87 Percentage: 5.95
Size: 511 Mflop/s: 2224.61 Percentage: 6.05
Size: 512 Mflop/s: 2284.85 Percentage: 6.21
Size: 639 Mflop/s: 2292.78 Percentage: 6.23
Size: 640 Mflop/s: 2264.7 Percentage: 6.15
Size: 767 Mflop/s: 2324.83 Percentage: 6.32
Size: 768 Mflop/s: 2363.92 Percentage: 6.42
Size: 769 Mflop/s: 2321.31 Percentage: 6.31
#Average percentage of Peak = 6.25811