Problem with ippiNormDiff_L1_16s in IPP 8.2

[Alexander_Karsakov]

Hello,

I have a strange result for ippiNormDiff_L1_16s_C1R function in IPP 8.2. Code for reproducing below:

short a[21*15] = { 10, 8, 7, 2, 3, 3, 6, 2, 2, 4, 7, 10, 8, 8, 6, 7, 5, 10, 7, 7, 3, 2, 3, 2, 6, 2, 6, 2, 9, 9, 9, 2, 4,
                10, 5, 5, 9, 2, 6, 7, 9, 9, 6, 3, 10, 7, 3, 9, 2, 10, 9, 10, 8, 8, 10, 5, 10, 9, 5, 4, 4, 2, 3, 5,
                4, 9, 5, 3, 7, 8, 2, 2, 7, 9, 10, 5, 5, 10, 5, 6, 5, 6, 7, 4, 10, 10, 3, 7, 2, 10, 7, 9, 2, 4, 9, 2,
                6, 6, 6, 6, 8, 3, 7, 10, 5, 5, 2, 4, 4, 7, 8, 10, 2, 8, 4, 2, 10, 3, 6, 4, 10, 9, 6, 10, 8, 5, 7, 8,
                10, 9, 3, 8, 4, 2, 3, 5, 3, 2, 3, 7, 4, 9, 10, 9, 8, 3, 4, 4, 9, 8, 7, 4, 10, 6, 2, 5, 10, 6, 3, 5,
                5, 10, 4, 9, 8, 5, 6, 6, 8, 7, 8, 6, 3, 4, 8, 8, 2, 6, 3, 3, 9, 10, 5, 8, 10, 9, 2, 8, 6, 7, 2, 10,
                7, 2, 8, 10, 5, 6, 4, 10, 9, 3, 6, 9, 2, 9, 10, 9, 8, 10, 5, 4, 10, 5, 10, 4, 4, 7, 10, 6, 8, 4, 7,
                5, 7, 5, 9, 4, 2, 2, 8, 5, 9, 3, 7, 2, 3, 2, 3, 8, 5, 9, 4, 4, 5, 4, 9, 7, 7, 8, 7, 5, 3, 8, 3, 6, 10,
                10, 3, 6, 9, 4, 4, 9, 4, 5, 9, 9, 4, 3, 8, 4, 3, 4, 6, 6, 10, 8, 5, 3, 8, 4, 6, 7, 9, 9, 9, 4, 2, 9,
                8, 5, 2, 10, 3, 4, 10, 5, 6, 3, 6, 10, 3, 5, 5, 5, 8, 6, 7, 8, 5, 3, 2, 2, 10 };

short b[21*15] = { 1358, -1177, 167, -1533, 62, -1416, -1002, 216, -844, 970, 1228, -663, 1503, 403,
                347, 621, -760, -920, 999, 1684, 809, 1507, -1383, -860, 1043, 1539, -1332, -437,
                54, -1033, 475, 183, -1401, 77, -963, 208, 1424, 21, 1295, 285, -69, -132, -204,
                884, 154, -1467, 1004, -1366, 1461, -190, 1564, 493, -238, 495, 853, -1316, 1057,
                -1534, 1107, 72, -284, -643, -1171, 509, 116, 1112, 177, 108, -402, 1422, -219,
                1163, -848, -502, 97, -1296, -498, 18, -185, 1438, -780, -1207, 1451, 1064, 876,
                -814, -483, 716, -1408, 1693, 1030, 228, -610, 486, 1076, 1540, 848, 1384, -831,
                -865, 452, -562, -611, 429, -740, 160, -8, 220, 1421, -1193, 860, -1422, 422,
                -1519, 1641, -1512, -251, 1151, -490, 150, 246, 1119, -822, -1057, 991, 979, 571,
                -692, 1399, -32, 1306, 1542, 1568, -1095, 1346, -736, -895, 1427, 1067, 488,
                -1430, 314, -85, -808, 672, 461, 960, 1117, 993, 729, -348, 1510, 1658, 1151, 570,
                -847, 704, 1219, -1103, -1176, 1590, -991, 893, -432, -194, -719, 811, -286, 1266,
                592, 895, -1321, -1531, 1051, -1352, -1355, -1109, -37, -1300, 875, 912, -1181,
                -339, 1431, 579, -1048, 752, 314, -1416, 1070, -1111, 911, -679, 1648, 340, 751,
                -1070, 758, 261, -1365, 681, 1505, -632, 1702, 705, -1001, 925, -736, 1432, -904,
                301, -1246, -448, 1177, 342, 1500, 1015, 767, 486, -812, -970, 999, -700, 571, 452,
                512, 802, -313, 1625, -184, 830, 1175, 894, 1233, -1317, 958, 98, 974, 1289, 947,
                -1449, -206, -109, 1332, -226, -190, 1247, 725, 227, -167, 1083, -1180, 789, -697,
                -683, -1440, -283, -1, -804, -217, 927, -90, 1321, 139, -166, 1015, -14, 1583,
                -682, 373, 1504, -1248, 1456, 215, 1105, -1338, -322, -389, 682, 694, 642, 1253,
                1306, 605, 468, 1296, 669, -417, -360, -479, 551, -476, -664, -301, -1476, 1603,
                229, -43, -1280, 1668, 1550, 556, 407, 853, 1516, -665, -1324, -366, 510, 1589,
                1248, 330, 1371, 1545, -147 };

IppiSize size = { 21, 15 };
Ipp64f res1;
ippiNormDiff_L1_16s_C1R(a, 21*sizeof(short), b, 21*sizeof(short), size, &res1);

double res2 = 0;
for (int i=0; i<21*15; i++)
    res2 += abs(a - b);

printf("IPP result = %f, my result = %f\n", res1, res2);

The result is:

IPP result = 263390.000000, my result = 263381.000000

Please look at this problem.

Best regards,

Alexander.

[Alexander_Karsakov]

My configuration:

• OS: Windows 7 HB
• Compiler: VC 17
• Processor: Intel Core i5 4570
• Intel® Parallel Studio XE Composer Edition for C++ Windows, package ID: w_ccompxe_2015.0.108.

[Igor_A_Intel · Accepted solution]

Hi Alexander,

thank you for your report/finding - I've reproduced the issue - it exists only in AVX2 code version - will be fixed in the next IPP version.

regards, Igor

[Alexander_Karsakov]

Hi Igor,

In IPPICV 8.2.1 we still have problems with ippiNormDiff_L1_16s_C1R. Now it failed another case (code for reproducing in attachment). Could you please take a look?

Best regards,

Alexander.

[Igor_A_Intel]

Hi Alexander,

got your message - will take a look and then answer - guess in 1 or 2 days.

regards, Igor

[Gennady_F_Intel]

hello!

the fix of the issue has been incorporated into IPP v.9.0 beta released recently.  You may evaluate the beta version and let us know how it will work. Here is the output which received with this version. 

ippIP AVX2 (l9 threaded) 9.0.0 Beta (r45861)
SSE    :Y
SSE2   :Y
SSE3   :Y
SSSE3  :Y
SSE41  :Y
SSE42  :Y
AVX    :Y
AVX2   :Y
----------
OS Enabled AVX :Y
AES            :Y
CLMUL          :Y
RDRAND         :Y
F16C           :Y
IPP = 29802861.000000, reference result = 29802861.000000
Press any key to continue . . .