1. Threshold Scan Inject various charges and histogram number of hits seen (occupancy plot) for each pixel Vcal (charge) from 0V to 200V (101 steps) For each Vcal 100 events triggered 32 mask steps Testbed: ROD21 in SR1 26 (4 PP0) modules connected to one ROD Mean and sigma of threshold by S-Curve fit RCE Meeting 10/26/09 M. Wittgen Page 1

2. S-Curve Log Likelihood Fit RCE Meeting 10/26/09 M. Wittgen Page 2 Number of electrons Number of hits

3. DSP Scan Engine RCE Meeting 10/26/09 M. Wittgen Page 3 Set Vcal

4. Hardware Limitation in ROD Dataflow Acquire data from modules (40MBit/s link per module) Trapping of events to SDSP 5 MBit/s serial port to send commands to modules Download histograms with SLAVE/MASTER/VME RCE Meeting 10/26/09 M. Wittgen Page 4

5. Leading Order Scan Steps Master Setup Mask Step Setup Parameter (Vcal for Example) Trigger Loop (100 events) Slave Processing Loop Fitting Corresponding slave steps for Setup Mask Step + Setup Parameter don't take much processing time RCE Meeting 10/26/09 M. Wittgen Page 5

6. Threshold Scan Timing Data processing (histogramming) enabled Trigger Loop MASTER: 305 s (101 x 32) Setup Mask Stage: 29 s (32) Setup Parameter: 142 s (32 x 101) Time in Processing SLAVE: 374 s Time in Fitting: 50 s No Data Processing on SLAVE - just counting words seen via DMA Trigger Loop: 170 s Time in Processing 128 s Time in Fit: 15 s (flatlined S-Curve) RCE Meeting 10/26/09 M. Wittgen Page 6

7. Scan Data Throughput Data Processing slows scan down by about a factor of 2 170 s vs 305 s for trigger loop Bytes seen on one SLAVE (7 modules): 6512 MByte / 170 s = 38 MByte/s 38 MByte/s / 7 modules = 5 MByte/s per module Consistent with a 40MBit/s stream + format overhead Amount of Data 6.5 GByte: 50 hits x 1 word per hit x 101 Vcal steps x 46080 pixel x 7 modules Occupancy histograms: 31 Mbyte = 101 Vcal Steps x 1 byte per pixel x 46080 pixel After fit: 3.7 Mbyte = 3 floats per pixel x 46080 pixels x 7 modules RCE Meeting 10/25/09 M. Wittgen Page 7

8. Serial Port Throughput Bottleneck: Creation of command bitstream + sending it to modules Lots of bit manipulation Setup Mask Stage: 180 KBit/s out of possible 5 MBit/s Dominated by bitstream setup but only takes 30 s Setup Parameter: 3.5 MBit/s out of 5 MBit/s Could be improved by using two serial ports? Twice as fast? 142 s / 2: could take of one min of scan time? RCE Meeting 10/26/09 M. Wittgen Page 8

9. Scan Time + Conclusion Performance limit 3 mins triggering + processing + 2 mins module configuration 5 mins theoretical limit for pure scan? Additionally 1 min for fitting Another limit is VME download From SLAVE to MASTER to SBC (last step through VME) At the moment 1 MByte/s out of 4 MByte/s possible Block transfer doesn't work (yet) Sequential download (one ROD after another) For threshold scan data (4 MByte per SLAVE) 4 s x 4 slaves x 16 RODs per crate = 256 s Additional PixLib (infrastructure) overhead RCE Meeting 10/26/09 M. Wittgen Page 9

10. DSP Code Characteristics 28K of C code Completely stand-alone (compiles with g++, TI cross compiler) 15 header files shared with PixLib Module configuration, primitives Pseudo-OO model (for example SLAVE Object – for communication with SLAVE from MASTER) RCE Meeting 10/26/09 M. Wittgen Page 10 typedef struct { int structType, id, fProc; int (*write)(void *slave, UINT32 addr, UINT32 *buffer, int count); int (*read)(void *slave, UINT32 addr, UINT32 *buffer, int count); int (*init)(void *slave); void (*proc)(void *slave); void (*start)(void *slave); void (*put)(void *slave, UINT32 addr, UINT32 word); UINT32 *(*sendCommand)(void *commandChannel, int id, void *buffer, int nWords, int fetchReply); UINT32 (*get)(void *slave, UINT32 addr); UINT32 *hpic, *hpia, *hpid, *hpidi; Xface *xface; /* points to a SlaveXface structure */ CommandChannel *commandChannel; RequestChannel *requestChannel; int (*sendPrimitive)(void *slave, void *primBuffer, int size); TextBuffer *textBuffer[N_TEXTBUFFERS]; UINT32 histogramTask; /* albeit a UINT32 for passing around, it is a pointer on the slave */ } Slave;

11. DSP Code Characteristics Most of the code can be reused Easily “converted” to pure OO/C++ Some code obsolete Task structures and queue Co-operative multitasking – only queuing feature really used Loop structure useful Very flexible – allows fast implementation of new scans DSPs operate independent (parallelization) Trend to implement loops on DSP rather than PixLib Important Bottleneck: Communication to SLAVE DSP from SBC only through MASTER Shared “address space” for SLAVE/MASTER tasks would be plus RCE Meeting 10/26/09 M. Wittgen Page 11