make check failure with PGI 12.8 with HDF5

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

I’ve been trying to build HDF5 (parallel with MVAPICH2 1.8) using PGI 12.8 and I keep encountering an error when I do a ‘make check’. The build runs just fine with no errors, but ‘make check’ dies with:

============================
Testing  hyperslab 
============================
 hyperslab  Test Log
============================
Test sizes:  SMALL MEDIUM
Testing hyperslab fill 11          variable hyperslab                  PASSED
Testing hyperslab fill 11x10       variable hyperslab                  PASSED
Testing hyperslab fill 3x5x5       variable hyperslab                  PASSED
Testing hyperslab fill 113         variable hyperslab                  PASSED
Testing hyperslab fill 15x11       variable hyperslab                  PASSED
Testing hyperslab fill 5x7x7       variable hyperslab                 *FAILED*
   at hyperslab.c:248 in test_fill()...
   acc != ref_value
   i=0, j=0, k=0, dx=1, dy=1, dz=1, fill=0
Reference array:
i=0:
000: 128 129 130 131 132 133 134
001: 135 136 137 138 139 140 141
002: 142 143 144 145 146 147 148
003: 149 150 151 152 153 154 155
004: 156 157 158 159 160 161 162
005: 163 164 165 166 167 168 169
006: 170 171 172 173 174 175 176

i=1:
000: 177 178 179 180 181 182 183
001: 184 185 186 187 188 189 190
002: 191 192 193 194 195 196 197
003: 198 199 200 201 202 203 204
004: 205 206 207 208 209 210 211
005: 212 213 214 215 216 217 218
006: 219 220 221 222 223 224 225

i=2:
000: 226 227 228 229 230 231 232
001: 233 234 235 236 237 238 239
002: 240 241 242 243 244 245 246
003: 247 248 249 250 251 252 253
004: 254 255   0   1   2   3   4
005:   5   6   7   8   9  10  11
006:  12  13  14  15  16  17  18

i=3:
000:  19  20  21  22  23  24  25
001:  26  27  28  29  30  31  32
002:  33  34  35  36  37  38  39
003:  40  41  42  43  44  45  46
004:  47  48  49  50  51  52  53
005:  54  55  56  57  58  59  60
006:  61  62  63  64  65  66  67

i=4:
000:  68  69  70  71  72  73  74
001:  75  76  77  78  79  80  81
002:  82  83  84  85  86  87  88
003:  89  90  91  92  93  94  95
004:  96  97  98  99 100 101 102
005: 103 104 105 106 107 108 109
006: 110 111 112 113 114 115 116


   Result is:
i=0:
000:   0 129 130 131 132 133 134
001: 135 136 137 138 139 140 141
002: 142 143 144 145 146 147 148
003: 149 150 151 152 153 154 155
004: 156 157 158 159 160 161 162
005: 163 164 165 166 167 168 169
006: 170 171 172 173 174 175 176

i=1:
000: 177 178 179 180 181 182 183
001: 184 185 186 187 188 189 190
002: 191 192 193 194 195 196 197
003: 198 199 200 201 202 203 204
004: 205 206 207 208 209 210 211
005: 212 213 214 215 216 217 218
006: 219 220 221 222 223 224 225

i=2:
000: 226 227 228 229 230 231 232
001: 233 234 235 236 237 238 239
002: 240 241 242 243 244 245 246
003: 247 248 249 250 251 252 253
004: 254 255   0   1   2   3   4
005:   5   6   7   8   9  10  11
006:  12  13  14  15  16  17  18

i=3:
000:  19  20  21  22  23  24  25
001:  26  27  28  29  30  31  32
002:  33  34  35  36  37  38  39
003:  40  41  42  43  44  45  46
004:  47  48  49  50  51  52  53
005:  54  55  56  57  58  59  60
006:  61  62  63  64  65  66  67

i=4:
000:  68  69  70  71  72  73  74
001:  75  76  77  78  79  80  81
002:  82  83  84  85  86  87  88
003:  89  90  91  92  93  94  95
004:  96  97  98  99 100 101 102
005: 103 104 105 106 107 108 109
006: 110 111 112 113 114 115 116

Testing hyperslab copy 11          variable source                     PASSED
Testing hyperslab copy 11          variable destination                PASSED
Testing hyperslab copy 11          sync source & dest                  PASSED
Testing hyperslab copy 179         variable source                     PASSED
Testing hyperslab copy 179         variable destination                PASSED
Testing hyperslab copy 179         sync source & dest                  PASSED
Testing hyperslab copy 11x10       variable source                     PASSED
Testing hyperslab copy 11x10       variable destination                PASSED
Testing hyperslab copy 11x10       sync source & dest                  PASSED
Testing hyperslab copy 13x19       variable source                     PASSED
Testing hyperslab copy 13x19       variable destination                PASSED
Testing hyperslab copy 13x19       sync source & dest                  PASSED
Testing hyperslab copy 73x67       variable source                     PASSED
Testing hyperslab copy 73x67       variable destination                PASSED
Testing hyperslab copy 73x67       sync source & dest                  PASSED
Testing hyperslab copy 3x5x5       variable source                     PASSED
Testing hyperslab copy 3x5x5       variable destination                PASSED
Testing hyperslab copy 3x5x5       sync source & dest                  PASSED
Testing hyperslab copy 7x9x5       variable source                     PASSED
Testing hyperslab copy 7x9x5       variable destination                PASSED
Testing hyperslab copy 7x9x5       sync source & dest                  PASSED
Testing multi-byte fill value                                          PASSED
Testing multi-byte fill value                                          PASSED
Testing endian conversion by stride                                    PASSED
Testing 2d transpose by stride    9x9d                                 PASSED
Testing 2d transpose by stride    3x11d                                PASSED
Testing endian conversion by stride                                    PASSED
Testing 2d transpose by stride 1200x1200d                              PASSED
Testing 2d transpose by stride  800x1800d                              PASSED
Testing image sampling   10x20   to    5x10                            PASSED
Testing image sampling    5x10   to   10x20                            PASSED
Testing image sampling  960x1280 to  480x640                           PASSED
Testing image sampling  480x640  to  960x1280                          PASSED
Testing array filling    1-9    elements                               PASSED
Testing array filling    9-257  elements                               PASSED
Testing array offset   13x  11x   7 elements                           PASSED
Testing array offset  347x 193x  71 elements                           PASSED
***** 1 HYPERSLAB TEST FAILED! *****
Command exited with non-zero status 1
0.55user 0.02system 0:00.58elapsed 99%CPU (0avgtext+0avgdata 157024maxresident)k
0inputs+0outputs (0major+23283minor)pagefaults 0swaps
make[4]: *** [hyperslab.chkexe_] Error 1

I’ve tested both hdf5-1.8.5-patch1 (which is our local operational version) and I just tried with hdf5-1.8.9 (which is the above). Both bomb out in ‘make check’ at that exact test with the same error in i=0, j=0, k=0, dx=1, dy=1, dz=1, fill=0.

The odd thing is that PGI 12.5 does not have this issue. If I set CC=‘mpicc -V12.5’ there is no problem and the code does a ‘make check’ just fine (well, there seems to be an issue with ph5diff, but that isn’t crucial for us).

Any suggestions on how to get around this? For edification, the configure used was:

          export PATH="$(prefix)/bin:$(PATH)" ;\
          export LDFLAGS="-lm" ;\
          ./configure --prefix=$(prefix) \
                      --includedir=$(prefix)/include/hdf5 \
                      --with-szlib=$(prefix)/include/szlib,$(prefix)/lib \
                      --with-zlib=$(prefix)/include/zlib,$(prefix)/lib \
                      --disable-shared --disable-cxx\
                      --enable-hl --enable-fortran --disable-sharedlib-rpath \
                      --enable-parallel \
                      CFLAGS="$(CFLAGS)" CC=$(NC_CC) FC=$(NC_FC) CXX=$(NC_CXX) )

where NC_CC=mpicc, NC_FC=mpif90, and NC_CXX=mpicxx and CFLAGS=-fPIC, and the prefix let’s the configure know where the prerequisite libraries are. Note: both szlib and zlib pass their ‘make check’.

Any ideas what changed between 12.5 and 12.8 that could cause this?

Thanks,
Matt

#2

Hi Matt,

This looks very similar to an issue that was reported to us last week (TPR#18920). What’s happening is that the “hyperslab” test is doing an accumulation into an unsigned int that wraps. When the loop that these variables are accumulated is optimized using either Unrolling or Vectorization, these variables get promoted to a 64-bit register for 4 operations. Because of this, the variable doesn’t wrap until it stored back to memory and hence gets a slightly different answer.

To work around the problem, replace “-fast” with “-fast -Mnovect -Mnounroll” to the “test” directory’s Makefile, or promote the variables “acc” and "ref_ " in “hyperslab.c” to “unsigned long”. This will allow the test to pass. Note that this error doesn’t seem to effect the HDF5 library itself, just the test.

I’ll add this info to TPR#18920.

Thanks,
Mat