SPRAC21A June   2016  – June 2019 OMAP-L132 , OMAP-L138 , TDA2E , TDA2EG-17 , TDA2HF , TDA2HG , TDA2HV , TDA2LF , TDA2P-ABZ , TDA2P-ACD , TDA2SA , TDA2SG , TDA2SX , TDA3LA , TDA3LX , TDA3MA , TDA3MD , TDA3MV

 

  1.   TDA2xx and TDA2ex Performance
    1.     Trademarks
    2. 1  SoC Overview
      1. 1.1 Introduction
      2. 1.2 Acronyms and Definitions
      3. 1.3 TDA2xx and TDA2ex System Interconnect
      4. 1.4 Traffic Regulation Within the Interconnect
        1. 1.4.1 Bandwidth Regulators
        2. 1.4.2 Bandwidth Limiters
        3. 1.4.3 Initiator Priority
      5. 1.5 TDA2xx and TDA2ex Memory Subsystem
        1. 1.5.1 Controller/PHY Timing Parameters
        2. 1.5.2 Class of Service
        3. 1.5.3 Prioritization Between DMM/SYS PORT or MPU Port to EMIF
      6. 1.6 TDA2xx and TDA2ex Measurement Operating Frequencies
      7. 1.7 System Instrumentation and Measurement Methodology
        1. 1.7.1 GP Timers
        2. 1.7.2 L3 Statistic Collectors
    3. 2  Cortex-A15
      1. 2.1 Level1 and Level2 Cache
      2. 2.2 MMU
      3. 2.3 Performance Control Mechanisms
        1. 2.3.1 Cortex-A15 Knobs
        2. 2.3.2 MMU Page Table Knobs
      4. 2.4 Cortex-A15 CPU Read and Write Performance
        1. 2.4.1 Cortex-A15 Functions
        2. 2.4.2 Setup Limitations
        3. 2.4.3 System Performance
          1. 2.4.3.1 Cortex-A15 Stand-Alone Memory Read, Write, Copy
          2. 2.4.3.2 Results
    4. 3  System Enhanced Direct Memory Access (System EDMA)
      1. 3.1 System EDMA Performance
        1. 3.1.1 System EDMA Read and Write
        2. 3.1.2 System EDMA Results
      2. 3.2 System EDMA Observations
    5. 4  DSP Subsystem EDMA
      1. 4.1 DSP Subsystem EDMA Performance
        1. 4.1.1 DSP Subsystem EDMA Read and Write
        2. 4.1.2 DSP Subsystem EDMA Results
      2. 4.2 DSP Subsystem EDMA Observations
    6. 5  Embedded Vision Engine (EVE) Subsystem EDMA
      1. 5.1 EVE EDMA Performance
        1. 5.1.1 EVE EDMA Read and Write
        2. 5.1.2 EVE EDMA Results
      2. 5.2 EVE EDMA Observations
    7. 6  DSP CPU
      1. 6.1 DSP CPU Performance
        1. 6.1.1 DSP CPU Read and Write
        2. 6.1.2 Code Setup
          1. 6.1.2.1 Pipeline Copy
          2. 6.1.2.2 Pipeline Read
          3. 6.1.2.3 Pipeline Write
          4. 6.1.2.4 L2 Stride-Jmp Copy
          5. 6.1.2.5 L2 Stride-Jmp Read
          6. 6.1.2.6 L2 Stride-Jmp Write
      2. 6.2 DSP CPU Observations
      3. 6.3 Summary
    8. 7  Cortex-M4 (IPU)
      1. 7.1 Cortex-M4 CPU Performance
        1. 7.1.1 Cortex-M4 CPU Read and Write
        2. 7.1.2 Code Setup
        3. 7.1.3 Cortex-M4 Functions
        4. 7.1.4 Setup Limitations
      2. 7.2 Cortex-M4 CPU Observations
        1. 7.2.1 Cache Disable
        2. 7.2.2 Cache Enable
      3. 7.3 Summary
    9. 8  USB IP
      1. 8.1 Overview
      2. 8.2 USB IP Performance
        1. 8.2.1 Test Setup
        2. 8.2.2 Results and Observations
        3. 8.2.3 Summary
    10. 9  PCIe IP
      1. 9.1 Overview
      2. 9.2 PCIe IP Performance
        1. 9.2.1 Test Setup
        2. 9.2.2 Results and Observations
    11. 10 IVA-HD IP
      1. 10.1 Overview
      2. 10.2 H.264 Decoder
        1. 10.2.1 Description
        2. 10.2.2 Test Setup
        3. 10.2.3 Test Results
      3. 10.3 MJPEG Decoder
        1. 10.3.1 Description
        2. 10.3.2 Test Setup
        3. 10.3.3 Test Results
    12. 11 MMC IP
      1. 11.1 MMC Read and Write Performance
        1. 11.1.1 Test Description
        2. 11.1.2 Test Results
      2. 11.2 Summary
    13. 12 SATA IP
      1. 12.1 SATA Read and Write Performance
        1. 12.1.1 Test Setup
        2. 12.1.2 Observations
          1. 12.1.2.1 RAW Performance
          2. 12.1.2.2 SDK Performance
      2. 12.2 Summary
    14. 13 GMAC IP
      1. 13.1 GMAC Receive/Transmit Performance
        1. 13.1.1 Test Setup
        2. 13.1.2 Test Description
          1. 13.1.2.1 CPPI Buffer Descriptors
        3. 13.1.3 Test Results
          1. 13.1.3.1 Receive/Transmit Mode (see )
          2. 13.1.3.2 Receive Only Mode (see )
          3. 13.1.3.3 Transmit Only Mode (see )
      2. 13.2 Summary
    15. 14 GPMC IP
      1. 14.1 GPMC Read and Write Performance
        1. 14.1.1 Test Setup
          1. 14.1.1.1 NAND Flash
          2. 14.1.1.2 NOR Flash
        2. 14.1.2 Test Description
          1. 14.1.2.1 Asynchronous NAND Flash Read/Write Using CPU Prefetch Mode
          2. 14.1.2.2 Asynchronous NOR Flash Single Read
          3. 14.1.2.3 Asynchronous NOR Flash Page Read
          4. 14.1.2.4 Asynchronous NOR Flash Single Write
        3. 14.1.3 Test Results
      2. 14.2 Summary
    16. 15 QSPI IP
      1. 15.1 QSPI Read and Write Performance
        1. 15.1.1 Test Setup
        2. 15.1.2 Test Results
        3. 15.1.3 Analysis
          1. 15.1.3.1 Theoretical Calculations
          2. 15.1.3.2 % Efficiency
      2. 15.2 QSPI XIP Code Execution Performance
      3. 15.3 Summary
    17. 16 Standard Benchmarks
      1. 16.1 Dhrystone
        1. 16.1.1 Cortex-A15 Tests and Results
        2. 16.1.2 Cortex-M4 Tests and Results
      2. 16.2 LMbench
        1. 16.2.1 LMbench Bandwidth
          1. 16.2.1.1 TDA2xx and TDA2ex Cortex-A15 LMbench Bandwidth Results
          2. 16.2.1.2 TDA2xx and TDA2ex Cortex-M4 LMBench Bandwidth Results
          3. 16.2.1.3 Analysis
        2. 16.2.2 LMbench Latency
          1. 16.2.2.1 TDA2xx and TDA2ex Cortex-A15 LMbench Latency Results
          2. 16.2.2.2 TDA2xx and TDA2ex Cortex-M4 LMbench Latency Results
          3. 16.2.2.3 Analysis
      3. 16.3 STREAM
        1. 16.3.1 TDA2xx and TDA2ex Cortex-A15 STREAM Benchmark Results
        2. 16.3.2 TDA2xx and TDA2ex Cortex-M4 STREAM Benchmark Results
    18. 17 Error Checking and Correction (ECC)
      1. 17.1 OCMC ECC Programming
      2. 17.2 EMIF ECC Programming
      3. 17.3 EMIF ECC Programming to Starterware Code Mapping
      4. 17.4 Careabouts of Using EMIF ECC
        1. 17.4.1 Restrictions Due to Non-Availability of Read Modify Write ECC Support in EMIF
          1. 17.4.1.1 Un-Cached CPU Access of EMIF
          2. 17.4.1.2 Cached CPU Access of EMIF
          3. 17.4.1.3 Non CPU Access of EMIF Memory
          4. 17.4.1.4 Debugger Access of EMIF via the Memory Browser/Watch Window
          5. 17.4.1.5 Software Breakpoints While Debugging
        2. 17.4.2 Compiler Optimization
        3. 17.4.3 Restrictions Due to i882 Errata
        4. 17.4.4 How to Find Who Caused the Unaligned Quanta Writes After the Interrupt
      5. 17.5 Impact of ECC on Performance
    19. 18 DDR3 Interleaved vs Non-Interleaved
      1. 18.1 Interleaved versus Non-Interleaved Setup
      2. 18.2 Impact of Interleaved vs Non-Interleaved DDR3 for a Single Initiator
      3. 18.3 Impact of Interleaved vs Non-Interleaved DDR3 for Multiple Initiators
    20. 19 DDR3 vs DDR2 Performance
      1. 19.1 Impact of DDR2 vs DDR3 for a Single Initiator
      2. 19.2 Impact of DDR2 vs DDR3 for Multiple Initiators
    21. 20 Boot Time Profile
      1. 20.1 ROM Boot Time Profile
      2. 20.2 System Boot Time Profile
    22. 21 L3 Statistics Collector Programming Model
    23. 22 Reference
  2.   Revision History

6.1.2.3 Pipeline Write

The C code for the pipeline write function is:

void pipeline_write(unsigned byte_cnt) { long long *restrict dst = (long long *)ext_buf[1]; unsigned int wrStartTime, wrStopTime; int i; _nassert((int)dst == 0); wrStartTime = CSL_tscRead(); #pragma UNROLL(2) for (i=0; i<byte_cnt/8; i++) { dst[i] = 0xDEADDEAD; } wrStopTime = CSL_tscRead(); WBINVALIDATE wrDuration = (float)(wrStopTime-wrStartTime)/(DSP_FREQ/1000); }

The analysis of the scheduled iteration is given out by the compiler as:

;*----------------------------------------------------------------------------* ;* SOFTWARE PIPELINE INFORMATION ;* ;* Loop found in file : ../pipeline_loop.c ;* Loop source line : 89 ;* Loop opening brace source line : 96 ;* Loop closing brace source line : 98 ;* Loop Unroll Multiple : 2x ;* Known Minimum Trip Count : 1 ;* Known Max Trip Count Factor : 1 ;* Loop Carried Dependency Bound(^) : 0 ;* Unpartitioned Resource Bound : 1 ;* Partitioned Resource Bound(*) : 1 ;* Resource Partition: ;* A-side B-side ;* .L units 0 0 ;* .S units 0 0 ;* .D units 1* 1* ;* .M units 0 0 ;* .X cross paths 0 0 ;* .T address paths 1* 1* ;* Long read paths 0 0 ;* Long write paths 0 0 ;* Logical ops (.LS) 0 0 (.L or .S unit) ;* Addition ops (.LSD) 0 0 (.L or .S or .D unit) ;* Bound(.L .S .LS) 0 0 ;* Bound(.L .S .D .LS .LSD) 1* 1* ;* ;* Searching for software pipeline schedule at ... ;* ii = 1 Schedule found with 2 iterations in parallel ;*----------------------------------------------------------------------------* ;* SETUP CODE ;* ;* MV B6,A3 ;* ADD 8,A3,A3 ;* MV A4,B4 ;* MV A5,B5 ;* ;* SINGLE SCHEDULED ITERATION ;* ;* $C$C94: ;* 0 STDW .D2T2 B5:B4,*B6++(16) ; |97| ;* || STDW .D1T1 A5:A4,*A3++(16) ; |97| ;* || SPBR $C$C94 ;* 1 NOP 1 ;* 2 ; BRANCHCC OCCURS {$C$C94} ; |89| ;*----------------------------------------------------------------------------*

The pipeline can be viewed as in Figure 22. It can be observed that this code would make sure the two store engines are occupied every cycle of the pipeline.

dsp_cpu_pipeline_write_software_sprac21.gifFigure 22. DSP CPU Pipeline Write Software Pipeline

The CGEM (C66x CorePac) L2 cache controller can get up to 4 L2 line allocations in flight. Each allocation brings in 128 bytes. The XMC can get an additional 8 pre-fetch requests (also for 128 bytes) in flight. In the best case, L2 + XMC can get 12 × 128 = 1.5K bytes worth of requests outstanding at once.

NOTE

In terms of bus requests, the L2 cache controller and XMC actually make 64 byte requests, so this is actually 24 64-byte requests in the best case.

In steady state, however, this should drop to 8 × 128 = 1K bytes total in-flight, because XMC only sends additional pre-fetches in response to pre-fetch hits. This is because, the XMC only sends new pre-fetches in two cases: (1) on recognizing a new stream, and (2) on getting hits to an existing stream. In steady state, there are no new streams, so you are only in case (2). In the "100% pre-fetch hit" steady state case, the L2 misses will all hit in the XMC pre-fetch buffer and stop there, and the only traffic leaving XMC will be additional pre-fetches. Thus, the total number of outstanding requests is limited to the total number of outstanding pre-fetches.

To get higher performance, you would need an optimized copy loop that can get 4 L2 misses pipelined up as much as possible, or if XMC pre-fetch is enabled, at least 4 pre-fetch streams active to maximize out the DSP Subsystem busses. In order to emulate this behavior, the following functions were defined that read or write the first 64-bit word of the L2 cache line of 128 bytes. Since they access only the first word of the L2 cache line, they are named L2 Stride-Jmp Copy, L2 Stride-Jmp Read, and L2 Stride-Jmp Write, respectively.