#!/usr/bin/env python # coding: utf-8 # # Pipeline Performance Mysteries # Consider setting power profiles, see `cpupower-gui`. # In[11]: get_ipython().system('LANG= cpupower frequency-info') # In[12]: get_ipython().system('rm -Rf tmp') get_ipython().system('mkdir -p tmp') # In[13]: get_ipython().run_cell_magic('writefile', 'tmp/pipeline-perf.c', '\n#include \n#include "timing.h"\n\n\nint main()\n{\n int result = 0;\n\n {\n int a = 0, b = 0;\n\n timestamp_type t1;\n get_timestamp(&t1);\n\n for (int ntrips = 0; ntrips < 1000; ++ntrips)\n for (int i = 0; i< 400*1000; ++i)\n {\n a += i;\n a += i;\n }\n\n timestamp_type t2;\n get_timestamp(&t2);\n\n printf("a, a: elapsed time %g s\\n",\n timestamp_diff_in_seconds(t1, t2));\n result += a+b;\n }\n\n return result;\n}\n') # In[14]: get_ipython().system('cd tmp; gcc -std=gnu99 -lrt -I.. -opipeline-perf pipeline-perf.c') get_ipython().system('tmp/pipeline-perf') # Check that the compiler didn't do anything unexpected: # In[15]: # !objdump --disassemble tmp/pipeline-perf # Come up with variants of this that exhibit various behaviors of the execution pipeline: # In[ ]: # In[16]: get_ipython().system('cd tmp; gcc -std=gnu99 -lrt -I.. -opipeline-perf pipeline-perf.c') get_ipython().system('tmp/pipeline-perf') # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # * Scroll down for solution # # In[17]: get_ipython().run_cell_magic('writefile', 'tmp/pipeline-perf.c', '\n#include \n#include "timing.h"\n\n\nint main()\n{\n int result = 0;\n\n {\n int a = 0, b = 0;\n\n timestamp_type t1;\n get_timestamp(&t1);\n\n for (int ntrips = 0; ntrips < 1000; ++ntrips)\n for (int i = 0; i< 1000*1000; ++i)\n {\n a += i;\n a += i;\n }\n\n timestamp_type t2;\n get_timestamp(&t2);\n\n printf("a, a: elapsed time %g s\\n",\n timestamp_diff_in_seconds(t1, t2));\n result += a+b;\n }\n\n {\n int a = 0, b = 0;\n\n timestamp_type t1;\n get_timestamp(&t1);\n\n for (int ntrips = 0; ntrips < 1000; ++ntrips)\n for (int i = 0; i< 1000*1000; ++i)\n {\n a += i;\n b += i;\n }\n\n timestamp_type t2;\n get_timestamp(&t2);\n\n printf("a, b: elapsed time %g s\\n",\n timestamp_diff_in_seconds(t1, t2));\n result += a+b;\n }\n\n {\n int a = 0, b = 0;\n\n timestamp_type t1;\n get_timestamp(&t1);\n\n for (int ntrips = 0; ntrips < 1000; ++ntrips)\n for (int i = 0; i< 250*1000; ++i)\n {\n a += i;\n a += i;\n \n a += i;\n a += i;\n \n a += i;\n a += i;\n \n a += i;\n a += i;\n \n }\n\n timestamp_type t2;\n get_timestamp(&t2);\n\n printf("a, a unrolled: elapsed time %g s\\n",\n timestamp_diff_in_seconds(t1, t2));\n result += a+b;\n }\n {\n int a = 0, b = 0;\n\n timestamp_type t1;\n get_timestamp(&t1);\n\n for (int ntrips = 0; ntrips < 1000; ++ntrips)\n for (int i = 0; i< 250*1000; ++i)\n {\n a += i;\n a += i;\n a += i;\n a += i;\n \n b += i;\n b += i;\n b += i;\n b += i;\n }\n\n timestamp_type t2;\n get_timestamp(&t2);\n\n printf("aa, bb unrolled: elapsed time %g s\\n",\n timestamp_diff_in_seconds(t1, t2));\n result += a+b;\n }\n\n {\n int a = 0, b = 0;\n\n timestamp_type t1;\n get_timestamp(&t1);\n\n for (int ntrips = 0; ntrips < 1000; ++ntrips)\n for (int i = 0; i< 250*1000; ++i)\n {\n a += i;\n b += i;\n\n a += i;\n b += i;\n\n a += i;\n b += i;\n\n a += i;\n b += i;\n }\n\n timestamp_type t2;\n get_timestamp(&t2);\n\n printf("a, b unrolled: elapsed time %g s\\n",\n timestamp_diff_in_seconds(t1, t2));\n result += a+b;\n }\n\n return result;\n}\n') # In[18]: get_ipython().system('cd tmp; gcc -std=gnu99 -lrt -I.. -opipeline-perf pipeline-perf.c') get_ipython().system('tmp/pipeline-perf') # In[ ]: