HW3: RC3 - Ampere measurements added

homework_3/analyse/RC3-a5ae5c6/Pending-PIDs-ampere

@@ -0,0 +1,68 @@
+[cchoutou@aristotle4 homework_3]$ ./hpc/submitJobs.sh
+Submitting: hpc/BitncV0Q20.sh
+Submitted batch job 1914925
+Submitting: hpc/BitncV0Q21.sh
+Submitted batch job 1914926
+Submitting: hpc/BitncV0Q22.sh
+Submitted batch job 1914927
+Submitting: hpc/BitncV0Q23.sh
+Submitted batch job 1914928
+Submitting: hpc/BitncV0Q24.sh
+Submitted batch job 1914929
+Submitting: hpc/BitncV0Q25.sh
+Submitted batch job 1914930
+Submitting: hpc/BitncV0Q26.sh
+Submitted batch job 1914931
+Submitting: hpc/BitncV0Q27.sh
+Submitted batch job 1914932
+Submitting: hpc/BitncV0Q28.sh
+Submitted batch job 1914933
+Submitting: hpc/BitncV0Q29.sh
+Submitted batch job 1914934
+Submitting: hpc/BitncV0Q30.sh
+Submitted batch job 1914935
+Submitting: hpc/BitncV1Q20.sh
+Submitted batch job 1914936
+Submitting: hpc/BitncV1Q21.sh
+Submitted batch job 1914937
+Submitting: hpc/BitncV1Q22.sh
+Submitted batch job 1914938
+Submitting: hpc/BitncV1Q23.sh
+Submitted batch job 1914939
+Submitting: hpc/BitncV1Q24.sh
+Submitted batch job 1914940
+Submitting: hpc/BitncV1Q25.sh
+Submitted batch job 1914941
+Submitting: hpc/BitncV1Q26.sh
+Submitted batch job 1914942
+Submitting: hpc/BitncV1Q27.sh
+Submitted batch job 1914943
+Submitting: hpc/BitncV1Q28.sh
+Submitted batch job 1914944
+Submitting: hpc/BitncV1Q29.sh
+Submitted batch job 1914945
+Submitting: hpc/BitncV1Q30.sh
+Submitted batch job 1914946
+Submitting: hpc/BitncV2Q20.sh
+Submitted batch job 1914947
+Submitting: hpc/BitncV2Q21.sh
+Submitted batch job 1914948
+Submitting: hpc/BitncV2Q22.sh
+Submitted batch job 1914949
+Submitting: hpc/BitncV2Q23.sh
+Submitted batch job 1914950
+Submitting: hpc/BitncV2Q24.sh
+Submitted batch job 1914951
+Submitting: hpc/BitncV2Q25.sh
+Submitted batch job 1914952
+Submitting: hpc/BitncV2Q26.sh
+Submitted batch job 1914953
+Submitting: hpc/BitncV2Q27.sh
+Submitted batch job 1914954
+Submitting: hpc/BitncV2Q28.sh
+Submitted batch job 1914955
+Submitting: hpc/BitncV2Q29.sh
+Submitted batch job 1914956
+Submitting: hpc/BitncV2Q30.sh
+Submitted batch job 1914957
+

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914925.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     1048576 (Q=20)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 8694 [usec]
+[Timing] Mem-xch : 5435 [usec]
+[Timing] Sorting : 3159 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914926.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     2097152 (Q=21)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 16 [msec]
+[Timing] Mem-xch : 12 [msec]
+[Timing] Sorting : 4093 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914927.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     4194304 (Q=22)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 53 [msec]
+[Timing] Mem-xch : 46 [msec]
+[Timing] Sorting : 6080 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914928.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     8388608 (Q=23)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 41 [msec]
+[Timing] Mem-xch : 27 [msec]
+[Timing] Sorting : 14 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914929.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     16777216 (Q=24)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 102 [msec]
+[Timing] Mem-xch : 72 [msec]
+[Timing] Sorting : 29 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914930.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     33554432 (Q=25)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 95 [msec]
+[Timing] Mem-xch : 36 [msec]
+[Timing] Sorting : 59 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914931.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     67108864 (Q=26)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 204 [msec]
+[Timing] Mem-xch : 84 [msec]
+[Timing] Sorting : 120 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914932.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     134217728 (Q=27)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 427 [msec]
+[Timing] Mem-xch : 172 [msec]
+[Timing] Sorting : 255 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914933.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     268435456 (Q=28)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 953 [msec]
+[Timing] Mem-xch : 396 [msec]
+[Timing] Sorting : 552 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914934.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     536870912 (Q=29)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 1859 [msec]
+[Timing] Mem-xch : 694 [msec]
+[Timing] Sorting : 1164 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914935.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V0
+[Log]: Array size:     1073741824 (Q=30)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 3769 [msec]
+[Timing] Mem-xch : 1316 [msec]
+[Timing] Sorting : 2452 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914936.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     1048576 (Q=20)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 3119 [usec]
+[Timing] Mem-xch : 1708 [usec]
+[Timing] Sorting : 1412 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914937.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     2097152 (Q=21)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 5227 [usec]
+[Timing] Mem-xch : 3052 [usec]
+[Timing] Sorting : 2200 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914938.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     4194304 (Q=22)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 9131 [usec]
+[Timing] Mem-xch : 5421 [usec]
+[Timing] Sorting : 3712 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914939.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     8388608 (Q=23)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 21 [msec]
+[Timing] Mem-xch : 12 [msec]
+[Timing] Sorting : 8485 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914940.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     16777216 (Q=24)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 34 [msec]
+[Timing] Mem-xch : 17 [msec]
+[Timing] Sorting : 17 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914941.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     33554432 (Q=25)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 70 [msec]
+[Timing] Mem-xch : 33 [msec]
+[Timing] Sorting : 37 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914942.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     67108864 (Q=26)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 144 [msec]
+[Timing] Mem-xch : 66 [msec]
+[Timing] Sorting : 79 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914943.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     134217728 (Q=27)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 329 [msec]
+[Timing] Mem-xch : 175 [msec]
+[Timing] Sorting : 153 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914944.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     268435456 (Q=28)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 763 [msec]
+[Timing] Mem-xch : 408 [msec]
+[Timing] Sorting : 354 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914945.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     536870912 (Q=29)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 1425 [msec]
+[Timing] Mem-xch : 685 [msec]
+[Timing] Sorting : 739 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914946.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V1
+[Log]: Array size:     1073741824 (Q=30)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 3201 [msec]
+[Timing] Mem-xch : 1626 [msec]
+[Timing] Sorting : 1555 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914947.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     1048576 (Q=20)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 3192 [usec]
+[Timing] Mem-xch : 1553 [usec]
+[Timing] Sorting : 1638 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914948.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     2097152 (Q=21)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 5811 [usec]
+[Timing] Mem-xch : 3228 [usec]
+[Timing] Sorting : 2629 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914949.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     4194304 (Q=22)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 12 [msec]
+[Timing] Mem-xch : 7923 [usec]
+[Timing] Sorting : 4624 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914950.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     8388608 (Q=23)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 18 [msec]
+[Timing] Mem-xch : 8502 [usec]
+[Timing] Sorting : 9900 [usec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914951.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     16777216 (Q=24)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 38 [msec]
+[Timing] Mem-xch : 17 [msec]
+[Timing] Sorting : 21 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914952.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     33554432 (Q=25)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 73 [msec]
+[Timing] Mem-xch : 28 [msec]
+[Timing] Sorting : 44 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914953.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     67108864 (Q=26)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 163 [msec]
+[Timing] Mem-xch : 80 [msec]
+[Timing] Sorting : 82 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914954.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     134217728 (Q=27)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 479 [msec]
+[Timing] Mem-xch : 301 [msec]
+[Timing] Sorting : 178 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914955.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     268435456 (Q=28)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 736 [msec]
+[Timing] Mem-xch : 328 [msec]
+[Timing] Sorting : 410 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914956.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     536870912 (Q=29)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 1595 [msec]
+[Timing] Mem-xch : 738 [msec]
+[Timing] Sorting : 857 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/analyse/RC3-a5ae5c6/ampere/slurm-1914957.out

@@ -0,0 +1,23 @@
+[Log]: Code version:   V2
+[Log]: Array size:     1073741824 (Q=30)
+[Log]: Repeated sorts: 7
+[Log]: GPU:            NVIDIA A100-SXM4-40GB
+[Log]: Block size:     512
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Log]: Initialize array ...  Done.
+[Log]: Start sorting    ...  Done.
+[Timing] Total   : 3237 [msec]
+[Timing] Mem-xch : 1443 [msec]
+[Timing] Sorting : 1793 [msec]
+[Validation] Results validation ... [PASSED] 

homework_3/src/bitonicsort.hpp

@@ -112,8 +112,8 @@ template <typename ValueT>
 __device__ void exchange(ValueT* data, threadId_t tid, threadId_t partner, bool keepSmall) {
     if (( keepSmall && (data[tid] > data[partner])) ||
         (!keepSmall && (data[tid] < data[partner])) ) {
-        ValueT temp = data[tid];
+        ValueT temp   = data[tid];
-        data[tid] = data[partner];
+        data[tid]     = data[partner];
         data[partner] = temp;
     }
 }
@@ -378,6 +378,7 @@ __global__ void interBlockStep(ValueT* data, size_t n, size_t step, size_t stage
     }
 }
 
+
 /*!
  * This is unrolled part of the bitonic double loop.
  *
@@ -399,42 +400,44 @@ __global__ void inBlockStep(ValueT* data, size_t n, size_t innerSteps, size_t st
      * Here we skip blocks every time (one for SizeToThreadsRatio = 2)
      * And we cache the neighbor block address indexes in local (shared) memory
      */
-    threadId_t gIdx0 = threadIdx.x + SizeToThreadsRatio * blockIdx.x * blockDim.x;
+    threadId_t gIdx = threadIdx.x + SizeToThreadsRatio * blockIdx.x * blockDim.x;
-    threadId_t lIdx0 = toLocal(gIdx0, blockDim.x);
+    threadId_t lIdx = toLocal(gIdx, blockDim.x);
 
-    if (gIdx0 + blockDim.x >= n)   // Boundary check
+    if (gIdx + blockDim.x >= n)   // Boundary check
         return;
 
     // Fetch to local memory the entire effective block size (2 positions for each thread)
-    shared_data[lIdx0]              = data[gIdx0];
+    shared_data[lIdx]              = data[gIdx];
-    shared_data[lIdx0 + blockDim.x] = data[gIdx0 + blockDim.x];
+    shared_data[lIdx + blockDim.x] = data[gIdx + blockDim.x];
     __syncthreads();
 
     for (size_t step = innerSteps + 1; step > 0; ) {
         --step;
 
-        // Init thread global and local indices
-        threadId_t gIdx = gIdx0;
-        threadId_t lIdx = lIdx0;
         // Find partner and keep-small configuration based on the global data positions
         threadId_t pIdx  = partner(gIdx, step);
         if (gIdx > pIdx) {
-            // Shift inside effective block
+            // Work on the right site
-            gIdx += blockDim.x; // global
+            bool keep = keepSmall(gIdx + blockDim.x, pIdx + blockDim.x, stage);
-            pIdx += blockDim.x;
-            lIdx += blockDim.x;	// local
-        }
-        bool keep = keepSmall(gIdx, pIdx, stage);
 
-        // Exchange data on local(shared) copy
+            // Exchange data on local(shared) copy
-        threadId_t lpIdx = toLocal(pIdx, blockDim.x);
+            threadId_t lpIdx  = toLocal(pIdx + blockDim.x, blockDim.x);
-        exchange(shared_data, lIdx, lpIdx, keep);
+            exchange(shared_data, lIdx + blockDim.x, lpIdx, keep);
+        }
+        else {
+            // Work on the left site
+            bool keep = keepSmall(gIdx, pIdx, stage);
+
+            // Exchange data on local(shared) copy
+            threadId_t lpIdx  = toLocal(pIdx, blockDim.x);
+            exchange(shared_data, lIdx, lpIdx, keep);
+        }
         __syncthreads();
     }
 
     // Write back to global memory (no sync here, there will be sync from host)
-    data[gIdx0]              = shared_data[lIdx0];
+    data[gIdx]              = shared_data[lIdx];
-    data[gIdx0 + blockDim.x] = shared_data[lIdx0 + blockDim.x];
+    data[gIdx + blockDim.x] = shared_data[lIdx + blockDim.x];
 }
 
 /*!
@@ -459,6 +462,59 @@ __global__ void prephase(ValueT* data, size_t n, size_t stages, size_t maxStages
      * Here we skip blocks every time (one for SizeToThreadsRatio = 2)
      * And we cache the neighbor block address indexes in local (shared) memory
      */
+    threadId_t gIdx = threadIdx.x + SizeToThreadsRatio * blockIdx.x * blockDim.x;
+    threadId_t lIdx = toLocal(gIdx, blockDim.x);
+
+    if (gIdx + blockDim.x >= n)   // Boundary check
+        return;
+
+    // Fetch to local memory the entire effective block size (2 positions for each thread)
+    shared_data[lIdx]              = data[gIdx];
+    shared_data[lIdx + blockDim.x] = data[gIdx + blockDim.x];
+    __syncthreads();
+
+    for (size_t stage = 1; (stage <= stages) && (stage <= maxStages); ++stage) {
+        for (size_t step = stage; step > 0; ) {
+            --step;
+
+            // Find partner and keep-small configuration based on the global data positions
+            threadId_t pIdx  = partner(gIdx, step);
+            if (gIdx > pIdx) {
+                // Work on the right site
+                bool keep = keepSmall(gIdx + blockDim.x, pIdx + blockDim.x, stage);
+
+                // Exchange data on local(shared) copy
+                threadId_t lpIdx  = toLocal(pIdx + blockDim.x, blockDim.x);
+                exchange(shared_data, lIdx + blockDim.x, lpIdx, keep);
+            }
+            else {
+                // Work on the left site
+                bool keep = keepSmall(gIdx, pIdx, stage);
+
+                // Exchange data on local(shared) copy
+                threadId_t lpIdx  = toLocal(pIdx, blockDim.x);
+                exchange(shared_data, lIdx, lpIdx, keep);
+            }
+            __syncthreads();
+        }
+    }
+
+    // Write back to global memory (no sync here, there will be sync from host)
+    data[gIdx]              = shared_data[lIdx];
+    data[gIdx + blockDim.x] = shared_data[lIdx + blockDim.x];
+
+#if 0
+    /*
+     * Idea:
+     * - Keep a register copy of data[gIdx0], and data[gIdx0 + blockDim.x]
+     * - Instead of exchange in shared_data, read in register the partner and exchange there.
+     * - Write back only if there was an exchange
+     *
+     * ^^
+     * Unfortunately this breaks sequential consistency and register values (lValve) does not match with share_data
+     * or even lValueR0 and lValueL0. Maybe there is something to do with register spilling (lValue keeps spill
+     * on local mem).
+     */
     threadId_t gIdx0 = threadIdx.x + SizeToThreadsRatio * blockIdx.x * blockDim.x;
     threadId_t lIdx0 = toLocal(gIdx0, blockDim.x);
 
@@ -466,29 +522,40 @@ __global__ void prephase(ValueT* data, size_t n, size_t stages, size_t maxStages
         return;
 
     // Fetch to local memory the entire effective block size (2 positions for each thread)
-    shared_data[lIdx0]              = data[gIdx0];
+    // also keep thread's init values (L and R) on register locations
-    shared_data[lIdx0 + blockDim.x] = data[gIdx0 + blockDim.x];
+    ValueT lValueL0                 = data[gIdx0];
+    ValueT lValueR0                 = data[gIdx0 + blockDim.x];
+    shared_data[lIdx0]              = lValueL0;
+    shared_data[lIdx0 + blockDim.x] = lValueR0;
     __syncthreads();
+
     for (size_t stage = 1; (stage <= stages) && (stage <= maxStages); ++stage) {
         for (size_t step = stage; step > 0; ) {
             --step;
 
-            // Init thread global and local indices
+            // Init thread global, local indices and active local register value
             threadId_t gIdx = gIdx0;
             threadId_t lIdx = lIdx0;
+            ValueT lValue   = lValueL0; // "Me" on the left side of effective block
+
             // Find partner and keep-small configuration based on the global data positions
             threadId_t pIdx  = partner(gIdx, step);
             if (gIdx > pIdx) {
                 // Shift inside effective block
-                gIdx += blockDim.x; // global
+                gIdx += blockDim.x;  // global
                 pIdx += blockDim.x;
-                lIdx += blockDim.x;	// local
+                lIdx += blockDim.x;  // local
+                lValue = lValueR0;   // The other me (the right side)
             }
             bool keep = keepSmall(gIdx, pIdx, stage);
 
             // Exchange data on local(shared) copy
             threadId_t lpIdx = toLocal(pIdx, blockDim.x);
-            exchange(shared_data, lIdx, lpIdx, keep);
+            ValueT pValue = shared_data[lpIdx];
+            if (exchangeVals(&lValue, &pValue, keep)) {
+                shared_data[lIdx]  = lValue;
+                shared_data[lpIdx] = pValue;
+            }
             __syncthreads();
         }
     }
@@ -496,8 +563,10 @@ __global__ void prephase(ValueT* data, size_t n, size_t stages, size_t maxStages
     // Write back to global memory (no sync here, there will be sync from host)
     data[gIdx0]              = shared_data[lIdx0];
     data[gIdx0 + blockDim.x] = shared_data[lIdx0 + blockDim.x];
+#endif
 }
 
+
 /*!
  * A CUDA version of the Bitonic sort algorithm.
  *

homework_3/src/config.h

@@ -16,8 +16,10 @@
 /*
  * Versioning:
  * - RC1: First version to test on HPC
+ * - RC2: A pre-phase added for v1 and v2
+ * - RC3:
  */
-static constexpr char version[] = "0.1";
+static constexpr char version[] = "0.2";
 
 /*
  * Defines for different version of the exercise