GCC 4.9 Linux LTO Optimizations Some link-time optimization compiler benchmarks by Michael Larabel for Phoronix.com of GCC 4.8.2 and 4.9.0 RC1. The Link-time optimization results with -flto didn't turn out to be as exciting as anticipated, so here they are for this short future article on phoronix. Benchmarks from an Intel Core i7 Haswell running Ubuntu Linux.
HTML result view exported from: https://openbenchmarking.org/result/1404126-PTS-GCC4849L62&grr&rdt .
GCC 4.9 Linux LTO Optimizations Processor Motherboard Chipset Memory Disk Graphics Audio Monitor Network OS Kernel Desktop Display Server Display Driver OpenGL Compiler File-System Screen Resolution GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO Intel Core i7-4770K @ 3.50GHz (8 Cores) ECS Z87H3-A2X EXTREME v1.0 Intel 4th Gen Core DRAM 16384MB 120GB Samsung SSD 840 ECS NVIDIA GeForce GTX 460 768MB (675/1804MHz) Realtek ALC1150 Samsung SyncMaster Realtek RTL8111/8168/8411 Ubuntu 14.04 3.13.0-22-generic (x86_64) Unity 7.2.0 X Server 1.15.0 NVIDIA 337.12 4.3.0 GCC 4.8.2 ext4 2560x1600 GCC 4.9.0 20140411 OpenBenchmarking.org Compiler Details - --enable-checking=release Processor Details - Scaling Governor: acpi-cpufreq ondemand
GCC 4.9 Linux LTO Optimizations apache: Static Web Page Serving hint: FLOAT nero2d: Total Time ffmpeg: H.264 HD To NTSC DV encode-mp3: WAV To MP3 encode-flac: WAV To FLAC smallpt: Global Illumination Renderer; 100 Samples open-porous-media: Upscale-Relperm ebizzy: Records/s c-ray: Total Time build-php: Time To Compile build-imagemagick: Time To Compile build-apache: Time To Compile himeno: Poisson Pressure Solver graphics-magick: Local Adaptive Thresholding graphics-magick: HWB Color Space graphics-magick: Resizing graphics-magick: Sharpen graphics-magick: Blur byte: Dhrystone 2 hpcc: G-Rand Access hpcc: EP-STREAM Triad hpcc: G-Ptrans hpcc: EP-DGEMM hpcc: G-Ffte hpcc: G-HPL GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 35904.53 367476347.00 453.45 13.99 12.42 4.82 24 100.58 42849 16.98 26.27 60.35 27.14 1810.10 97 216 198 140 170 35908267.37 0.00979 1.08895 1.26591 6.57694 2.11922 50.05477 35521.51 367633749.83 458.86 13.80 13.24 24 100.42 42698 16.98 93.29 124.51 44.37 1813.55 97 218 199 141 171 34717223.53 0.00980 1.08612 1.26956 6.63749 2.12967 49.78783 35953.73 373674384.83 13.80 10.87 3.70 42950 17.09 27.05 60.59 27.88 1828.15 102 214 199 140 166 36694838.47 0.00980 1.08175 1.26904 6.72611 2.11338 50.08117 35817.51 371516290.66 13.75 10.88 3.72 42565 17.08 95.42 129.91 45.74 1825.41 102 213 198 140 165 33544122.87 0.00989 1.08126 1.26850 6.63834 2.13555 49.96230 OpenBenchmarking.org
Apache Benchmark Static Web Page Serving OpenBenchmarking.org Requests Per Second, More Is Better Apache Benchmark 2.4.7 Static Web Page Serving GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 8K 16K 24K 32K 40K SE +/- 136.25, N = 3 SE +/- 75.16, N = 3 SE +/- 76.61, N = 3 SE +/- 55.60, N = 3 35904.53 35521.51 35953.73 35817.51 -flto -flto 1. (CC) gcc options: -shared -fPIC -pthread -O3 -march=native
Hierarchical INTegration Test: FLOAT OpenBenchmarking.org QUIPs, More Is Better Hierarchical INTegration 1.0 Test: FLOAT GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 80M 160M 240M 320M 400M SE +/- 103032.27, N = 3 SE +/- 129752.70, N = 3 SE +/- 362392.17, N = 3 SE +/- 994638.92, N = 3 367476347.00 367633749.83 373674384.83 371516290.66 -flto -flto 1. (CC) gcc options: -O3 -march=native -lm
Open FMM Nero2D Total Time OpenBenchmarking.org Seconds, Fewer Is Better Open FMM Nero2D 2.0.2 Total Time GCC 4.8.2 - Stock GCC 4.8.2 - LTO 100 200 300 400 500 453.45 458.86 -flto 1. (CXX) g++ options: -O3 -march=native -lfftw3 -llapack -lf77blas -latlas -lm
FFmpeg H.264 HD To NTSC DV OpenBenchmarking.org Seconds, Fewer Is Better FFmpeg 2.1.1 H.264 HD To NTSC DV GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 4 8 12 16 20 SE +/- 0.03, N = 3 SE +/- 0.07, N = 3 SE +/- 0.01, N = 3 SE +/- 0.06, N = 3 13.99 13.80 13.80 13.75 -flto -flto 1. (CC) gcc options: -lavdevice -lavfilter -lavformat -lavcodec -lswresample -lswscale -lavutil -ldl -lasound -lSDL -lm -pthread -O3 -march=native -std=c99 -fomit-frame-pointer -fno-math-errno -fno-signed-zeros -fno-tree-vectorize -MMD -MF -MT
LAME MP3 Encoding WAV To MP3 OpenBenchmarking.org Seconds, Fewer Is Better LAME MP3 Encoding 3.99.3 WAV To MP3 GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 3 6 9 12 15 SE +/- 0.04, N = 5 SE +/- 0.05, N = 5 SE +/- 0.01, N = 5 SE +/- 0.01, N = 5 12.42 13.24 10.87 10.88 -flto -flto 1. (CC) gcc options: -pipe -O3 -march=native -lm
FLAC Audio Encoding WAV To FLAC OpenBenchmarking.org Seconds, Fewer Is Better FLAC Audio Encoding 1.3.0 WAV To FLAC GCC 4.8.2 - Stock GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 1.0845 2.169 3.2535 4.338 5.4225 SE +/- 0.06, N = 7 SE +/- 0.09, N = 10 SE +/- 0.09, N = 10 4.82 3.70 3.72 -flto 1. (CXX) g++ options: -O3 -march=native -fvisibility=hidden -logg -lm
Smallpt Global Illumination Renderer; 100 Samples OpenBenchmarking.org Seconds, Fewer Is Better Smallpt 1.0 Global Illumination Renderer; 100 Samples GCC 4.8.2 - Stock GCC 4.8.2 - LTO 6 12 18 24 30 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 24 24 -flto 1. (CXX) g++ options: -fopenmp -O3 -march=native
Open Porous Media OPM Benchmark: Upscale-Relperm OpenBenchmarking.org Seconds, Fewer Is Better Open Porous Media 2013-11-26 OPM Benchmark: Upscale-Relperm GCC 4.8.2 - Stock GCC 4.8.2 - LTO 20 40 60 80 100 SE +/- 0.17, N = 3 SE +/- 0.38, N = 3 100.58 100.42 1. (F9X) gfortran options: -rdynamic
ebizzy Records/s OpenBenchmarking.org Seconds, More Is Better ebizzy 0.3 Records/s GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 9K 18K 27K 36K 45K SE +/- 109.85, N = 3 SE +/- 104.13, N = 3 SE +/- 44.08, N = 3 SE +/- 33.07, N = 3 42849 42698 42950 42565 -flto -flto 1. (CC) gcc options: -pthread -lpthread -O3 -march=native
C-Ray Total Time OpenBenchmarking.org Seconds, Fewer Is Better C-Ray 1.1 Total Time GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 4 8 12 16 20 SE +/- 0.01, N = 3 SE +/- 0.00, N = 3 SE +/- 0.01, N = 3 SE +/- 0.01, N = 3 16.98 16.98 17.09 17.08 -flto -flto 1. (CC) gcc options: -lm -lpthread -O3 -march=native
Timed PHP Compilation Time To Compile OpenBenchmarking.org Seconds, Fewer Is Better Timed PHP Compilation 5.2.9 Time To Compile GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 20 40 60 80 100 SE +/- 0.08, N = 3 SE +/- 1.16, N = 3 SE +/- 0.02, N = 3 SE +/- 0.14, N = 3 26.27 93.29 27.05 95.42 -flto -flto 1. (CC) gcc options: -O3 -march=native -pedantic -ldl -lz -lm
Timed ImageMagick Compilation Time To Compile OpenBenchmarking.org Seconds, Fewer Is Better Timed ImageMagick Compilation 6.8.1-10 Time To Compile GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 30 60 90 120 150 SE +/- 0.36, N = 3 SE +/- 0.30, N = 3 SE +/- 0.04, N = 3 SE +/- 0.32, N = 3 60.35 124.51 60.59 129.91
Timed Apache Compilation Time To Compile OpenBenchmarking.org Seconds, Fewer Is Better Timed Apache Compilation 2.4.7 Time To Compile GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 10 20 30 40 50 SE +/- 0.19, N = 3 SE +/- 0.07, N = 3 SE +/- 0.23, N = 3 SE +/- 0.05, N = 3 27.14 44.37 27.88 45.74
Himeno Benchmark Poisson Pressure Solver OpenBenchmarking.org MFLOPS, More Is Better Himeno Benchmark 3.0 Poisson Pressure Solver GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 400 800 1200 1600 2000 SE +/- 3.93, N = 3 SE +/- 0.48, N = 3 SE +/- 1.31, N = 3 SE +/- 0.63, N = 3 1810.10 1813.55 1828.15 1825.41 -flto -flto 1. (CC) gcc options: -O3 -march=native
GraphicsMagick Operation: Local Adaptive Thresholding OpenBenchmarking.org Iterations Per Minute, More Is Better GraphicsMagick 1.3.19 Operation: Local Adaptive Thresholding GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 20 40 60 80 100 SE +/- 0.33, N = 3 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 97 97 102 102 -flto -flto 1. (CC) gcc options: -std=gnu99 -fopenmp -O3 -march=native -pthread -ljbig -lwebp -ljpeg -lXext -lX11 -llzma -lxml2 -lz -lm -lpthread
GraphicsMagick Operation: HWB Color Space OpenBenchmarking.org Iterations Per Minute, More Is Better GraphicsMagick 1.3.19 Operation: HWB Color Space GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 50 100 150 200 250 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 SE +/- 0.33, N = 3 SE +/- 0.00, N = 3 216 218 214 213 -flto -flto 1. (CC) gcc options: -std=gnu99 -fopenmp -O3 -march=native -pthread -ljbig -lwebp -ljpeg -lXext -lX11 -llzma -lxml2 -lz -lm -lpthread
GraphicsMagick Operation: Resizing OpenBenchmarking.org Iterations Per Minute, More Is Better GraphicsMagick 1.3.19 Operation: Resizing GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 40 80 120 160 200 SE +/- 0.33, N = 3 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 SE +/- 0.33, N = 3 198 199 199 198 -flto -flto 1. (CC) gcc options: -std=gnu99 -fopenmp -O3 -march=native -pthread -ljbig -lwebp -ljpeg -lXext -lX11 -llzma -lxml2 -lz -lm -lpthread
GraphicsMagick Operation: Sharpen OpenBenchmarking.org Iterations Per Minute, More Is Better GraphicsMagick 1.3.19 Operation: Sharpen GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 30 60 90 120 150 SE +/- 0.33, N = 3 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 SE +/- 0.00, N = 3 140 141 140 140 -flto -flto 1. (CC) gcc options: -std=gnu99 -fopenmp -O3 -march=native -pthread -ljbig -lwebp -ljpeg -lXext -lX11 -llzma -lxml2 -lz -lm -lpthread
GraphicsMagick Operation: Blur OpenBenchmarking.org Iterations Per Minute, More Is Better GraphicsMagick 1.3.19 Operation: Blur GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 40 80 120 160 200 SE +/- 0.33, N = 3 SE +/- 0.33, N = 3 SE +/- 0.33, N = 3 SE +/- 0.33, N = 3 170 171 166 165 -flto -flto 1. (CC) gcc options: -std=gnu99 -fopenmp -O3 -march=native -pthread -ljbig -lwebp -ljpeg -lXext -lX11 -llzma -lxml2 -lz -lm -lpthread
BYTE Unix Benchmark Computational Test: Dhrystone 2 OpenBenchmarking.org LPS, More Is Better BYTE Unix Benchmark 3.6 Computational Test: Dhrystone 2 GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 8M 16M 24M 32M 40M SE +/- 13215.25, N = 3 SE +/- 9630.07, N = 3 SE +/- 165665.98, N = 3 SE +/- 32494.09, N = 3 35908267.37 34717223.53 36694838.47 33544122.87 -flto -flto 1. (CC) gcc options: -O3 -march=native
HPC Challenge Test / Class: G-Random Access OpenBenchmarking.org GUP/s, More Is Better HPC Challenge 1.4.3 Test / Class: G-Random Access GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 0.0022 0.0044 0.0066 0.0088 0.011 SE +/- 0.00008, N = 3 SE +/- 0.00001, N = 3 SE +/- 0.00013, N = 3 SE +/- 0.00003, N = 3 0.00979 0.00980 0.00980 0.00989 -flto -flto 1. (CC) gcc options: -lblas -lm -pthread -lmpi -ldl -lhwloc -fomit-frame-pointer -O3 -march=native -funroll-loops 2. BLAS + Open MPI 1.6.5
HPC Challenge Test / Class: EP-STREAM Triad OpenBenchmarking.org GB/s, More Is Better HPC Challenge 1.4.3 Test / Class: EP-STREAM Triad GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 0.245 0.49 0.735 0.98 1.225 SE +/- 0.00488, N = 3 SE +/- 0.00405, N = 3 SE +/- 0.00379, N = 3 SE +/- 0.00342, N = 3 1.08895 1.08612 1.08175 1.08126 -flto -flto 1. (CC) gcc options: -lblas -lm -pthread -lmpi -ldl -lhwloc -fomit-frame-pointer -O3 -march=native -funroll-loops 2. BLAS + Open MPI 1.6.5
HPC Challenge Test / Class: G-Ptrans OpenBenchmarking.org GB/s, More Is Better HPC Challenge 1.4.3 Test / Class: G-Ptrans GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 0.2857 0.5714 0.8571 1.1428 1.4285 SE +/- 0.00051, N = 3 SE +/- 0.00112, N = 3 SE +/- 0.00329, N = 3 SE +/- 0.00220, N = 3 1.26591 1.26956 1.26904 1.26850 -flto -flto 1. (CC) gcc options: -lblas -lm -pthread -lmpi -ldl -lhwloc -fomit-frame-pointer -O3 -march=native -funroll-loops 2. BLAS + Open MPI 1.6.5
HPC Challenge Test / Class: EP-DGEMM OpenBenchmarking.org GFLOPS, More Is Better HPC Challenge 1.4.3 Test / Class: EP-DGEMM GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 2 4 6 8 10 SE +/- 0.04345, N = 3 SE +/- 0.05029, N = 3 SE +/- 0.05569, N = 3 SE +/- 0.06945, N = 3 6.57694 6.63749 6.72611 6.63834 -flto -flto 1. (CC) gcc options: -lblas -lm -pthread -lmpi -ldl -lhwloc -fomit-frame-pointer -O3 -march=native -funroll-loops 2. BLAS + Open MPI 1.6.5
HPC Challenge Test / Class: G-Ffte OpenBenchmarking.org GFLOPS, More Is Better HPC Challenge 1.4.3 Test / Class: G-Ffte GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 0.4805 0.961 1.4415 1.922 2.4025 SE +/- 0.00495, N = 3 SE +/- 0.00327, N = 3 SE +/- 0.00571, N = 3 SE +/- 0.01396, N = 3 2.11922 2.12967 2.11338 2.13555 -flto -flto 1. (CC) gcc options: -lblas -lm -pthread -lmpi -ldl -lhwloc -fomit-frame-pointer -O3 -march=native -funroll-loops 2. BLAS + Open MPI 1.6.5
HPC Challenge Test / Class: G-HPL OpenBenchmarking.org GFLOPS, More Is Better HPC Challenge 1.4.3 Test / Class: G-HPL GCC 4.8.2 - Stock GCC 4.8.2 - LTO GCC 4.9.0 RC1 - Stock GCC 4.9.0 RC1 - LTO 11 22 33 44 55 SE +/- 0.08, N = 3 SE +/- 0.24, N = 3 SE +/- 0.05, N = 3 SE +/- 0.11, N = 3 50.05 49.79 50.08 49.96 -flto -flto 1. (CC) gcc options: -lblas -lm -pthread -lmpi -ldl -lhwloc -fomit-frame-pointer -O3 -march=native -funroll-loops 2. BLAS + Open MPI 1.6.5
Phoronix Test Suite v10.8.5