🐛 [BUG] - Performance Gap for Gather-Scatter Kernel

[radelja]

💿 OS

RHEL 8

🏗️ Architecture

x86_64

📝 Description

There is a performance gap between versions 1.1 and 2.0 of Spatter for the OpenMP gather-scatter kernel. This has been mentioned in the wiki, and this gap remains despite changes to the OpenMP kernels to align the disassembled instructions of Spatter 2.0 with version 1.1. The bandwidth reported by Spatter 2.0 is nearly half the bandwidth of version 1.1 on Cascade Lake and Sandy Bridge.

🔄 Reproduction steps

1. Go to project directory
2. Run CMake to build OpenMP backend with cmake -DUSE_OPENMP=1 -B build_openmp -S . && make -j$(nproc) -C build_openmp
3. Run Spatter with a gather-scatter pattern using -gUNIFORM:8:1 -uUNIFORM:8:1 -ksg -l16777216 -v2 -t12

💬 Logs

Running Spatter version 1.1
Compiler: GNU ver. 8.5.0
Backend: OpenMP
Aggregate Results? NO

Run Configurations
[ {'id': 0, 'kernel': 'sg', 'pattern': [], 'pattern-gather': [0, 1, 2, 3, 4, 5, 6, 7], 'pattern-scatter': [0, 1, 2, 3, 4, 5, 6, 7], 'delta': 8, 'delta-gather': 8, 'delta-scatter': 8, 'count': 16777216, 'wrap': 1, 'threads': 12} ]

config         bytes          time(s)        bw(MB/s)       
0              2147483648     0.0396006      54228.6