Skip to content

Latest commit

 

History

History
111 lines (94 loc) · 3.87 KB

README.md

File metadata and controls

111 lines (94 loc) · 3.87 KB
Raw

Scripts

Check build README on how to generate logs and add profiling

Analysis of instrucion log and trace

Script: ./run_analysis.py

Barebones approach
Run instruction log analysis

python3 run_analysis.py -i ../src/vector_mac_vm_uint8_inst_profiler.json
../sw/baremetal/vector_mac/vm_uint8.dump

Run trace analysis

python3 run_analysis.py -t ../src/vector_mac_vm_uint8_trace.bin

Most common approach
Run instruction log analysis with highlight

HL="lw,lh,lb,lhu,lbu,sw,sh,sb bne,beq,blt,bge,bgeu,bltu jal,jalr"
python3 run_analysis.py -i ../src/vector_mac_vm_uint8_inst_profiler.json --highlight $HL

Run trace analysis with annotated symbols and highlight

python3 run_analysis.py -t ../src/vector_mac_vm_uint8_trace.bin --dasm ../sw/baremetal/vector_mac/vm_uint8.dump --highlight $HL

For long traces, increase limit with --pc_time_series_limit 120000
Optionally, save symbols found in dasm for later annotation: --save_symbols
Filter out only part of the program by specifying program counter begin/end with: --pc_begin 0x80000094 --pc_end 0x800000ec

With --dasm, script reports symbols and number of executed instruction for each symbol (i.e. function)

Symbols found in ../sw/baremetal/vector_mac/vm_uint8.dump:
0x01CC - 0x02AC: main (10906)
0x01A8 - 0x01C8: __mulsi3 (43264)
0x0128 - 0x01A4: write_uart0 (0)
0x00F4 - 0x0124: fail (0)
0x00C4 - 0x00F0: pass (7)
0x00A4 - 0x00C0: write_csr_status (8)
0x0060 - 0x00A0: write_mismatch (0)
0x0048 - 0x005C: set_c (0)
0x0040 - 0x0044: asm_add (0)
0x002C - 0x003C: test_end (5)
0x0028 - 0x0028: done_bss (1)
0x0018 - 0x0024: clear_bss (257)
0x0000 - 0x0014: _start (6)

Both instruction log and trace analysis can accept:

  • --silent to suppress window pop-up
  • --save_png or --save_svg to save chart to disk
  • --save_csv to save source data formatted as csv to disk

For all options:

python3 run_analysis.py -h

Annotate execution log

Script: ./annotate_exec_log.py
Requires that symbols have been saved in during trace analysis

python3 annotate_exec_log.py -l ../src/vector_mac_vm_uint8_exec.log -s ../sw/baremetal/vector_mac/vm_uint8_symbols.json

HW performance estimates

Script: perf_est.py
HW config file: hw_perf_metrics.json

python3 perf_est.py ../src/vector_mac_vm_uint8_inst_profiler.json hw_perf_metrics.json

Config file, based on the ama-riscv implementation

{
    "cpu_frequency_mhz" : 125,
    "pipeline_latency" : 5,
    "branch_resolution" : 1,
    "jump_resolution" : 1,
    "memory_response" : 0,
    "mispredict_penalty" : 1,
    "prediction_resolution" : 0
}

pipeline_latency - number of clock cycles between reset being deasserted and the first instruction being committed
"branch_resolution" : 1 - branch would cause 1 clock bubble
"jump_resolution" : 1 - jump would cause 1 clock bubble
"memory_response" : 0 - DMEM responds in the next cycle
"mispredict_penalty" : 1 - one instruction is flushed on misprediction (since there is one cycle delay for branch resolution)
"prediction_resolution" : 0 - prediction is available in the next cycle

Output:

vector_mac_vm_uint8_inst_profiler.json
Branches total: 15537 out of 54454 total instructions (28.5% branches)
    Taken: 10030, Forwards: 2833, Backwards: 7197
    Not taken: 5507, Forwards: 4416, Backwards: 1091
    Predicted: 11613, Mispredicted: 3924, Accuracy: 74.7%
    Cycles: Original/With prediction: 72113/60500 (11613 cycles saved)
Potential app speedup: 16.1%
Estimated HW performance at 125MHz with 72113 cycles executed: CPI=1.32, exec time=576.9us, MIPS=94.4
Estimated HW performance at 125MHz with 60500 cycles executed: CPI=1.11, exec time=484.0us, MIPS=112.5