.results.sh 8.0 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190
  1. #!/bin/bash
  2. eval $nodes_;
  3. eval $cores_;
  4. eval $mpi_proc_;
  5. eval $threads_;
  6. eval $testcase_;
  7. eval $n_pts_;
  8. eval $m_;
  9. eval $q_;
  10. eval $tol_;
  11. eval $depth_;
  12. eval $unif_;
  13. eval $adap_;
  14. eval $max_time_;
  15. eval $fname_;
  16. eval $args_;
  17. export RESULT_FNAME=${RESULT_DIR}/${RESULT_FNAME}
  18. rm -f ${RESULT_FNAME}
  19. echo "#########################################################################################################" | tee -a ${RESULT_FNAME}
  20. echo "# CPU + MIC Results : Time (FLOP/s) [CPU Only, CPU+MIC, CPU+MIC (async)] #" | tee -a ${RESULT_FNAME}
  21. echo "#########################################################################################################" | tee -a ${RESULT_FNAME}
  22. ################### Input Parameter Fields ###################
  23. export PARAMSTR="FMM Kernel name;Order of multipole expansions;Order of Chebyshev polynomials;Maximum Tree Depth;Chebyshev Tolerance"
  24. IFS=';' read -ra PARAMARR <<< "$PARAMSTR"
  25. Nparam=${#PARAMARR[@]};
  26. export PARAMHEADERSTR="kernel;m;q;max_d;tol"
  27. IFS=';' read -ra PARAMHEADERARR <<< "$PARAMHEADERSTR"
  28. ################### Time (FLOP/s) Fields ###################
  29. export COLSTR="RunFMM;UpwardPass;ReduceBcast;DownwardPass;U-List;V-List;W-List;X-List;D2H_Wait:Trg;D2D;D2T"
  30. IFS=';' read -ra COLARR <<< "$COLSTR"
  31. N=${#COLARR[@]};
  32. ################### Error Fields ###################
  33. export ERRSTR="Maximum Relative Error \[Input\];Relative L2 Error \[Input\];Relative L2 Error \[Output\];Maximum Relative Error \[Output\];Relative L2 Error \[OutputGrad\];Maximum Relative Error \[OutputGrad\]"
  34. IFS=';' read -ra ERRARR <<< "$ERRSTR"
  35. Nerr=${#ERRARR[@]};
  36. export ERRHEADERSTR="Linf(f);L2(f);L2(u);Linf(u);L2(grad_u);Linf(grad_u)"
  37. IFS=';' read -ra ERRHEADERARR <<< "$ERRHEADERSTR"
  38. ################### Print Column Headers ###################
  39. printf "%$((16+$(($Nparam+4))*11+$N*18+$Nerr*14))s\n" |tr " " "=" | tee -a ${RESULT_FNAME} #=================================================
  40. #-----------------------------------------------------------
  41. for (( i=0; i<$Nparam; i++ )) ; do
  42. printf "%10s " "${PARAMHEADERARR[i]}" | tee -a ${RESULT_FNAME}
  43. done;
  44. printf " |" | tee -a ${RESULT_FNAME}
  45. #-----------------------------------------------------------
  46. HEADER_FORMAT="%10s "
  47. printf "${HEADER_FORMAT}" "MPI_PROC" | tee -a ${RESULT_FNAME}
  48. printf "${HEADER_FORMAT}" "THREADS" | tee -a ${RESULT_FNAME}
  49. printf "${HEADER_FORMAT}" "NODES" | tee -a ${RESULT_FNAME}
  50. printf "${HEADER_FORMAT}" "OCT/NODE" | tee -a ${RESULT_FNAME}
  51. printf " |" | tee -a ${RESULT_FNAME}
  52. #-----------------------------------------------------------
  53. for (( i=0; i<$N; i++ )) ; do
  54. printf " %17s" "${COLARR[i]}" | tee -a ${RESULT_FNAME}
  55. done;
  56. printf " |" | tee -a ${RESULT_FNAME}
  57. #-----------------------------------------------------------
  58. for (( i=0; i<$Nerr; i++ )) ; do
  59. printf " %13s" "${ERRHEADERARR[i]}" | tee -a ${RESULT_FNAME}
  60. done;
  61. printf " |\n" | tee -a ${RESULT_FNAME}
  62. #-----------------------------------------------------------
  63. printf "%$((16+$(($Nparam+4))*11+$N*18+$Nerr*14))s\n" |tr " " "=" | tee -a ${RESULT_FNAME} #=================================================
  64. #===========================================================
  65. ################### Loop over all runs ###################
  66. for (( l=0; l<${#nodes[@]}; l++ )) ; do
  67. ( # Begin parallel subshell
  68. RESULT_FNAME=${RESULT_FNAME}_${l};
  69. # File name.
  70. FNAME_NOMIC=${RESULT_DIR}/$(basename ${EXEC})_${fname[l]}.out;
  71. FNAME_MIC=${RESULT_DIR}/$(basename ${EXEC})_mic_${fname[l]}.out;
  72. FNAME_ASYNC=${RESULT_DIR}/$(basename ${EXEC})_async_${fname[l]}.out;
  73. subrow_cnt=0;
  74. for (( k=0; k<3; k++ )) ; do
  75. if [ $k -eq 0 ] ; then
  76. FNAME=${FNAME_NOMIC};
  77. fi
  78. if [ $k -eq 1 ] ; then
  79. FNAME=${FNAME_MIC};
  80. fi
  81. if [ $k -eq 2 ] ; then
  82. FNAME=${FNAME_ASYNC};
  83. fi
  84. if [ ! -f ${FNAME} ] ; then
  85. #echo >> ${RESULT_FNAME}
  86. continue;
  87. fi;
  88. subrow_cnt=$(( $subrow_cnt + 1 ))
  89. ######################### Parse Data #################################
  90. # Parse Data: Parameters
  91. for (( i=0; i<$Nparam; i++ )) ; do
  92. x="${PARAMARR[i]}"
  93. PARAM[i]="$(grep -hir "$x" ${FNAME} | tail -n 1 | rev | cut -d ' ' -f 1 | rev)";
  94. done
  95. #---------------------------------------------------------------------
  96. # Parse Data: Leaf Count
  97. NODE_LIST="$(grep -hir 'Leaf Nodes:' ${FNAME} | tail -n 1 | cut -d " " -f 3-)"
  98. NODES=0;
  99. for i in ${NODE_LIST} ; do
  100. NODES=$(( $NODES + $i ));
  101. done;
  102. #---------------------------------------------------------------------
  103. # Parse Data: Time, Flop, Flop/s
  104. for (( i=0; i<$N; i++ )) ; do
  105. x="${COLARR[i]}"
  106. T_MAX[i]="$(grep -hir "$x " ${FNAME} | tail -n 1 | tr -s ' ' | rev | cut -d ' ' -f 10 | rev)"
  107. if [ "${T_MAX[i]}" == "" ]; then continue; fi;
  108. FP_AVG[i]="$(grep -hir "$x " ${FNAME_NOMIC} | tail -n 1 | tr -s ' ' | rev | cut -d ' ' -f 8 | rev)"
  109. FLOPS[i]=$(echo "scale=10;${FP_AVG[i]}/(${T_MAX[i]}+0.0001)" | bc 2> /dev/null)
  110. if [ "${FLOPS[i]}" != "" ] && [ -f ${FNAME_MIC} ] && [ -f ${FNAME_ASYNC} ] && [ -f ${FNAME_NOMIC} ] ; then
  111. T_MAX_NOASYNC[i]="$(grep -hir "$x " ${FNAME_MIC} | tail -n 1 | tr -s ' ' | rev | cut -d ' ' -f 10 | rev)"
  112. if [ "${T_MAX_NOASYNC[i]}" == "" ]; then continue; fi;
  113. compare_result1=$(echo "${T_MAX[i]}<0.5*${T_MAX_NOASYNC[i]}" | bc)
  114. compare_result2=$(echo "${T_MAX[i]}<0.01" | bc)
  115. if [ ${compare_result1} -eq 1 ] && [ ${compare_result2} -eq 1 ] ; then
  116. FLOPS[i]=$(echo "scale=10;${FP_AVG[i]}/${T_MAX_NOASYNC[i]}" | bc 2> /dev/null)
  117. fi
  118. fi
  119. FLOPS[i]=$(echo "scale=10;${FLOPS[i]}*${mpi_proc[l]}/${nodes[l]}" | bc 2> /dev/null)
  120. done
  121. #---------------------------------------------------------------------
  122. # Parse Data: Error
  123. for (( i=0; i<$Nerr; i++ )) ; do
  124. x="${ERRARR[i]}"
  125. ERR[i]="$(grep -hir "$x" ${FNAME} | tail -n 1 | rev | cut -d ' ' -f 1 | rev)";
  126. done
  127. #=====================================================================
  128. ######################### Print Data #################################
  129. PARAM_FORMAT="%10s "
  130. for (( i=0; i<$Nparam; i++ )) ; do
  131. printf "${PARAM_FORMAT}" "${PARAM[i]}" >> ${RESULT_FNAME}
  132. done;
  133. printf " |" >> ${RESULT_FNAME}
  134. #---------------------------------------------------------------------
  135. printf "${PARAM_FORMAT}" "${mpi_proc[l]}" >> ${RESULT_FNAME}
  136. printf "${PARAM_FORMAT}" "${threads[l]}" >> ${RESULT_FNAME}
  137. printf "${PARAM_FORMAT}" "${nodes[l]}" >> ${RESULT_FNAME}
  138. printf "${PARAM_FORMAT}" "$((${NODES}/${nodes[l]}))" >> ${RESULT_FNAME}
  139. printf " |" >> ${RESULT_FNAME}
  140. #---------------------------------------------------------------------
  141. TIMING_FORMAT=" %9.3f (%5.1f)"
  142. for (( i=0; i<$N; i++ )) ; do
  143. printf "${TIMING_FORMAT}" "${T_MAX[i]}" "${FLOPS[i]}" >> ${RESULT_FNAME}
  144. done;
  145. printf " |" >> ${RESULT_FNAME}
  146. #---------------------------------------------------------------------
  147. ERR_FORMAT=" %1.2e"
  148. for (( i=0; i<$Nerr; i++ )) ; do
  149. printf "${ERR_FORMAT}" "${ERR[i]}" >> ${RESULT_FNAME}
  150. done;
  151. printf " |\n" >> ${RESULT_FNAME}
  152. #=====================================================================
  153. done
  154. if [[ $l == $(( ${#nodes[@]}-1 )) ]] || [ "${nodes[l]}" == ":" ]; then
  155. printf "%$((16+$(($Nparam+4))*11+$N*18+$Nerr*14))s\n" |tr " " "=" >> ${RESULT_FNAME} #=================================================
  156. elif [[ $subrow_cnt > 1 ]]; then
  157. printf "%$((16+$(($Nparam+4))*11+$N*18+$Nerr*14))s\n" |tr " " "-" >> ${RESULT_FNAME} #-------------------------------------------------
  158. fi
  159. )& # End parallel subshell
  160. # Combine results
  161. if (( ($l+1) % 10 == 0 )) || [[ $l == $(( ${#nodes[@]}-1 )) ]] ; then
  162. wait;
  163. for (( i=0; i<${#nodes[@]}; i++ )) ; do
  164. RESULT_FNAME_=${RESULT_FNAME}_${i};
  165. if [ -f ${RESULT_FNAME_} ] ; then
  166. cat ${RESULT_FNAME_} | tee -a ${RESULT_FNAME};
  167. rm ${RESULT_FNAME_};
  168. fi;
  169. done
  170. fi
  171. done