]>
Commit | Line | Data |
---|---|---|
1 | #ifndef ALIHLTTPCCAGPUCONFIG_H | |
2 | #define ALIHLTTPCCAGPUCONFIG_H | |
3 | ||
4 | //GPU Run Configuration | |
5 | ||
6 | //#define FERMI | |
7 | ||
8 | #ifdef FERMI | |
9 | #define HLTCA_GPU_BLOCK_COUNT_CONSTRUCTOR_MULTIPLIER 2 | |
10 | #define HLTCA_GPU_BLOCK_COUNT_SELECTOR_MULTIPLIER 3 | |
11 | #define HLTCA_GPU_THREAD_COUNT 256 | |
12 | #define HLTCA_GPU_THREAD_COUNT_CONSTRUCTOR 256 | |
13 | #define HLTCA_GPU_THREAD_COUNT_SELECTOR 256 | |
14 | #define HLTCA_GPU_THREAD_COUNT_FINDER 256 | |
15 | #else | |
16 | #define HLTCA_GPU_BLOCK_COUNT_CONSTRUCTOR_MULTIPLIER 1 | |
17 | #define HLTCA_GPU_BLOCK_COUNT_SELECTOR_MULTIPLIER 1 | |
18 | #define HLTCA_GPU_THREAD_COUNT 256 | |
19 | #define HLTCA_GPU_THREAD_COUNT_CONSTRUCTOR 256 | |
20 | #define HLTCA_GPU_THREAD_COUNT_SELECTOR 256 | |
21 | #define HLTCA_GPU_THREAD_COUNT_FINDER 256 | |
22 | #endif | |
23 | ||
24 | #define HLTCA_GPU_DEFAULT_HELPER_THREADS 2 //Number of helper threads to speed up initialization/output | |
25 | ||
26 | //GPU Parameters | |
27 | #define HLTCA_GPU_WARP_SIZE 32 | |
28 | #define HLTCA_GPU_REGS 64 | |
29 | ||
30 | //#define HLTCA_GPU_MERGER //Use GPU Merger | |
31 | ||
32 | //Detector Parameters | |
33 | #define HLTCA_ROW_COUNT 159 | |
34 | ||
35 | #define HLTCA_GPU_ROWALIGNMENT uint4 //Align Row Hits and Grid | |
36 | #define HLTCA_GPU_ROWCOPY int //must not be bigger than row alignment!!! | |
37 | ||
38 | #define HLTCA_GPU_SCHED_ROW_STEP 32 //Amount of Rows to process in one step before rescheduling | |
39 | #define HLTCA_GPU_SCHED_FIXED_START //Assign each GPU thread a start tracklet to start with instead of using the scheduler to obtain start tracklet | |
40 | //#define HLTCA_GPU_SCHED_FIXED_SLICE //Make each Multiprocessor on the GPU work only on a single slice during tracklet construction | |
41 | #define HLTCA_GPU_RESCHED //Use dynamic tracklet scheduling | |
42 | ||
43 | #define HLTCA_GPU_ALTERNATIVE_SCHEDULER //Use alternative scheduling algorithm (makes upper 4 options obsolete) | |
44 | #define HLTCA_GPU_ALTSCHED_STEPSIZE 80 //Number of rows to process in between of rescheduling | |
45 | #define HLTCA_GPU_ALTSCHED_MIN_THREADS 32 //Reschedule if less than n threads are active | |
46 | #define HLTCA_GPU_ALTERNATIVE_SCHEDULER_SIMPLE //Use simple version of alternative scheduler | |
47 | ||
48 | #ifndef FERMI | |
49 | #define HLTCA_GPU_TEXTURE_FETCH //Fetch data through texture cache | |
50 | #define HLTCA_GPU_TEXTURE_FETCHa //Fetch also in Neighbours Finder | |
51 | #endif | |
52 | ||
53 | //#define HLTCA_GPU_TRACKLET_CONSTRUCTOR_DO_PROFILE //Output Profiling Data for Tracklet Constructor Tracklet Scheduling | |
54 | //#define HLTCA_GPU_TIME_PROFILE //Output Time Profiling Data for asynchronous DMA transfer | |
55 | #define BITWISE_COMPATIBLE_DEBUG_OUTPUT //Make Debug Output of CPU and GPU bitwise compatible for comparison, also enable SORT_DUMPDATA! | |
56 | #define HLTCA_GPU_SORT_DUMPDATA //Sort Start Hits etc before dumping to file | |
57 | ||
58 | #define HLTCA_GPU_TRACKLET_SELECTOR_HITS_REG_SIZE 12 | |
59 | #define HLTCA_GPU_TRACKLET_SELECTOR_SLICE_COUNT 3 //Currently must be smaller than avaiable MultiProcessors on GPU or will result in wrong results | |
60 | ||
61 | #define HLTCA_GPU_MAX_TRACKLETS 12288 //Max Number of Tracklets that can be processed by GPU Tracker, Should be divisible by 16 at least | |
62 | #define HLTCA_GPU_MAX_TRACKS 1536 //Max number of Tracks that can be processd by GPU Tracker | |
63 | ||
64 | //#define HLTCA_GPU_EMULATION_SINGLE_TRACKLET 1313 //Run Tracklet constructor on on single Tracklet in Device Emulation Mode | |
65 | //#define HLTCA_GPU_EMULATION_DEBUG_TRACKLET 1313 | |
66 | ||
67 | //#define HLTCA_GPU_DEFAULT_MAX_SLICE_COUNT 12 | |
68 | ||
69 | #define HLTCA_GPU_TRACKER_CONSTANT_MEM 65000 //Amount of Constant Memory to reserve | |
70 | ||
71 | #define HLTCA_GPU_TRACKER_OBJECT_MEMORY 1024 * 1024 //Total amount of Memory to reserve for GPU Tracker Objects | |
72 | #define HLTCA_GPU_ROWS_MEMORY 1024 * 1024 //Total amount of Memory to reserve for GPU Row Parameters | |
73 | #define HLTCA_GPU_COMMON_MEMORY 1024 * 1024 //Total amount of Memory to reserve for CommomMemoryStruct on GPU | |
74 | #define HLTCA_GPU_SLICE_DATA_MEMORY 6 * 1024 * 1024 //Amount of Slice Data Memory to reserve per Slice on GPU | |
75 | #define HLTCA_GPU_GLOBAL_MEMORY 13 * 1024 * 1024 //Amount of global temporary Memory to reserve per Slice on GPU | |
76 | #define HLTCA_GPU_TRACKS_MEMORY 2 * 1024 * 1024 //Amount of Memory to reserve for Final Tracks per Slice on Host as Page Locked Memory | |
77 | ||
78 | //Make sure options do not interfere | |
79 | ||
80 | #ifndef HLTCA_GPUCODE | |
81 | //No texture fetch for CPU Tracker | |
82 | #ifdef HLTCA_GPU_TEXTURE_FETCH | |
83 | #undef HLTCA_GPU_TEXTURE_FETCH | |
84 | #endif | |
85 | #ifdef HLTCA_GPU_TEXTURE_FETCHa | |
86 | #undef HLTCA_GPU_TEXTURE_FETCHa | |
87 | #endif | |
88 | ||
89 | //Do not cache Row Hits during Tracklet selection in Registers for CPU Tracker | |
90 | #undef HLTCA_GPU_TRACKLET_SELECTOR_HITS_REG_SIZE | |
91 | #define HLTCA_GPU_TRACKLET_SELECTOR_HITS_REG_SIZE 0 | |
92 | #else | |
93 | //Sort start hits for GPU tracker | |
94 | #define HLTCA_GPU_SORT_STARTHITS | |
95 | #endif | |
96 | ||
97 | //Error Codes for GPU Tracker | |
98 | #define HLTCA_GPU_ERROR_NONE 0 | |
99 | #define HLTCA_GPU_ERROR_ROWBLOCK_TRACKLET_OVERFLOW 1 | |
100 | #define HLTCA_GPU_ERROR_TRACKLET_OVERFLOW 2 | |
101 | #define HLTCA_GPU_ERROR_TRACK_OVERFLOW 3 | |
102 | #define HLTCA_GPU_ERROR_SCHEDULE_COLLISION 4 | |
103 | #define HLTCA_GPU_ERROR_WRONG_ROW 5 | |
104 | ||
105 | #endif | |
106 |