This is a cpu tool for benchmarking the floating-points peak performance. Now it supports linux and x86-64 platform. It can automatically sense the local SIMD ISAs while compiling.
build x64 version:
./build_x64.sh
build arm64 version:
./build_arm64.sh
benchmark:
./cpufp --thread_pool=[xxx] --idle_time=yyy
clean:
./clean.sh
xxx indicates that all the cores defined by xxx will be used for benchmarking(by affinity setting). For example, [0,3,5-8,13-15].
idle_time is the interval time(sec) between two benchmarks, default 0.
| ISA | Data Type | Description |
|---|---|---|
| SSE | fp32 | Before Sandy Bridge |
| SSE2 | fp64 | Before Sandy Bridge |
| AVX | fp32/fp64 | From Sandy Bridge |
| FMA | fp32/fp64 | From Haswell/Zen |
| AVX512f | fp32/fp64 | From Skylake X/Zen4 |
| AVX512_VNNI | int8/int16 | From IceLake |
| AVX_VNNI | int8/int16 | From Alder Lake |
| AVX512_FP16 | fp16 | From Intel Sapphire Rapids |
| AVX512_BF16 | bf16 | From AMD Zen4 |
| AVX_VNNI_INT8 | int8 | Unknown |
| ISA | Data Type | Description |
|---|---|---|
| asimd | fp32/fp64 | From Cortex-A57/A53 |
| asimd_hp | fp16 | From Cortex-A75/A55 |
| asimd_dp | int8 | From Cortex-A75/A55 |
| bf16 | bf16 | From Cortex-X2/A710/A510 |
| i8mm | int8 | From Cortex-X2/A710/A510 |
For single core:
$ ./cpufp --thread_pool=[0] Number Threads: 1 Thread Pool Binding: 0 -------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | FMA | FMA(f32,f32,f32) | 151.84 GFLOPS | | FMA | FMA(f64,f64,f64) | 75.702 GFLOPS | | AVX | ADD(MUL(f32,f32),f32) | 150.86 GFLOPS | | AVX | ADD(MUL(f64,f64),f64) | 75.476 GFLOPS | | SSE | ADD(MUL(f32,f32),f32) | 75.452 GFLOPS | | SSE2 | ADD(MUL(f64,f64),f64) | 37.737 GFLOPS | --------------------------------------------------------------
For multi-cores:
$ ./cpufp --thread_pool=[0,2,4,6,8,10,12,14] Number Threads: 8 Thread Pool Binding: 0 2 4 6 8 10 12 14 -------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | FMA | FMA(f32,f32,f32) | 1057.8 GFLOPS | | FMA | FMA(f64,f64,f64) | 534.37 GFLOPS | | AVX | ADD(MUL(f32,f32),f32) | 1037.6 GFLOPS | | AVX | ADD(MUL(f64,f64),f64) | 516.21 GFLOPS | | SSE | ADD(MUL(f32,f32),f32) | 518.32 GFLOPS | | SSE2 | ADD(MUL(f64,f64),f64) | 258.92 GFLOPS | --------------------------------------------------------------
For single core:
$ ./cpufp --thread_pool=[0] Number Threads: 1 Thread Pool Binding: 0 -------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | AVX_VNNI | DP4A(s32,u8,s8) | 108.51 GOPS | | AVX_VNNI | DP2A(s32,s16,s16) | 54.244 GOPS | | FMA | FMA(f32,f32,f32) | 54.247 GFLOPS | | FMA | FMA(f64,f64,f64) | 27.128 GFLOPS | | AVX | ADD(MUL(f32,f32),f32) | 27.128 GFLOPS | | AVX | ADD(MUL(f64,f64),f64) | 13.564 GFLOPS | | SSE | ADD(MUL(f32,f32),f32) | 27.126 GFLOPS | | SSE2 | ADD(MUL(f64,f64),f64) | 13.563 GFLOPS | --------------------------------------------------------------
For multi_cores:
$ ./cpufp --thread_pool=[0-3] Number Threads: 4 Thread Pool Binding: 0 1 2 3 -------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | AVX_VNNI | DP4A(s32,u8,s8) | 369.66 GOPS | | AVX_VNNI | DP2A(s32,s16,s16) | 185.09 GOPS | | FMA | FMA(f32,f32,f32) | 185.08 GFLOPS | | FMA | FMA(f64,f64,f64) | 92.55 GFLOPS | | AVX | ADD(MUL(f32,f32),f32) | 92.546 GFLOPS | | AVX | ADD(MUL(f64,f64),f64) | 46.269 GFLOPS | | SSE | ADD(MUL(f32,f32),f32) | 92.546 GFLOPS | | SSE2 | ADD(MUL(f64,f64),f64) | 46.27 GFLOPS | --------------------------------------------------------------
For single core:
$ ./cpufp --thread_pool=[0] Number Threads: 1 Thread Pool Binding: 0 ------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd | fmla.vs(f32,f32,f32) | 11.958 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 11.958 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 5.9792 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 5.9792 GFLOPS | -------------------------------------------------------------
For multi_cores:
$ ./cpufp --thread_pool=[0-3] Number Threads: 4 Thread Pool Binding: 0 1 2 3 ------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd | fmla.vs(f32,f32,f32) | 47.883 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 47.88 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 23.933 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 23.943 GFLOPS | -------------------------------------------------------------
For single core:
$ ./cpufp --thread_pool=[0] Number Threads: 1 Thread Pool Binding: 0 ---------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd_dp | dp4a.vs(s32,s8,s8) | 153.48 GOPS | | asimd_dp | dp4a.vv(s32,s8,s8) | 153.48 GOPS | | asimd_dp | dp4a.vs(u32,u8,u8) | 153.47 GOPS | | asimd_dp | dp4a.vv(u32,u8,u8) | 153.48 GOPS | | asimd_hp | fmla.vs(fp16,fp16,fp16) | 76.738 GFLOPS | | asimd_hp | fmla.vv(fp16,fp16,fp16) | 76.738 GFLOPS | | asimd | fmla.vs(f32,f32,f32) | 38.369 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 38.369 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 19.185 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 19.185 GFLOPS | ----------------------------------------------------------------
For multi_cores:
$ ./cpufp --thread_pool=[0-3] Number Threads: 4 Thread Pool Binding: 0 1 2 3 ---------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd_dp | dp4a.vs(s32,s8,s8) | 613.79 GOPS | | asimd_dp | dp4a.vv(s32,s8,s8) | 614.02 GOPS | | asimd_dp | dp4a.vs(u32,u8,u8) | 613.98 GOPS | | asimd_dp | dp4a.vv(u32,u8,u8) | 613.99 GOPS | | asimd_hp | fmla.vs(fp16,fp16,fp16) | 306.88 GFLOPS | | asimd_hp | fmla.vv(fp16,fp16,fp16) | 306.98 GFLOPS | | asimd | fmla.vs(f32,f32,f32) | 153.48 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 153.5 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 74.513 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 76.751 GFLOPS | ----------------------------------------------------------------
For single core(Cortex-A55):
$ ./cpufp --thread_pool=[0] Number Threads: 1 Thread Pool Binding: 0 ---------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd_dp | dp4a.vs(s32,s8,s8) | 58.305 GOPS | | asimd_dp | dp4a.vv(s32,s8,s8) | 58.311 GOPS | | asimd_dp | dp4a.vs(u32,u8,u8) | 58.313 GOPS | | asimd_dp | dp4a.vv(u32,u8,u8) | 58.311 GOPS | | asimd_hp | fmla.vs(fp16,fp16,fp16) | 29.156 GFLOPS | | asimd_hp | fmla.vv(fp16,fp16,fp16) | 29.156 GFLOPS | | asimd | fmla.vs(f32,f32,f32) | 14.579 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 14.577 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 7.2891 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 7.2834 GFLOPS | ----------------------------------------------------------------
For multi_cores(Cortex-A55):
$ ./cpufp --thread_pool=[0-3] Number Threads: 4 Thread Pool Binding: 0 1 2 3 ---------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd_dp | dp4a.vs(s32,s8,s8) | 232.58 GOPS | | asimd_dp | dp4a.vv(s32,s8,s8) | 232.46 GOPS | | asimd_dp | dp4a.vs(u32,u8,u8) | 232.59 GOPS | | asimd_dp | dp4a.vv(u32,u8,u8) | 232.54 GOPS | | asimd_hp | fmla.vs(fp16,fp16,fp16) | 116.29 GFLOPS | | asimd_hp | fmla.vv(fp16,fp16,fp16) | 116.28 GFLOPS | | asimd | fmla.vs(f32,f32,f32) | 58.145 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 58.14 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 29.072 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 29.07 GFLOPS | ----------------------------------------------------------------
For single core(Cortex-A76):
$ ./cpufp --thread_pool=[4] Number Threads: 1 Thread Pool Binding: 4 ---------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd_dp | dp4a.vs(s32,s8,s8) | 151.74 GOPS | | asimd_dp | dp4a.vv(s32,s8,s8) | 151.75 GOPS | | asimd_dp | dp4a.vs(u32,u8,u8) | 151.75 GOPS | | asimd_dp | dp4a.vv(u32,u8,u8) | 151.74 GOPS | | asimd_hp | fmla.vs(fp16,fp16,fp16) | 75.862 GFLOPS | | asimd_hp | fmla.vv(fp16,fp16,fp16) | 75.862 GFLOPS | | asimd | fmla.vs(f32,f32,f32) | 37.927 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 37.925 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 18.961 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 18.961 GFLOPS | ----------------------------------------------------------------
For multi_cores(Cortex-A76):
$ ./cpufp --thread_pool=[4-7] Number Threads: 4 Thread Pool Binding: 4 5 6 7 ---------------------------------------------------------------- | Instruction Set | Core Computation | Peak Performance | | asimd_dp | dp4a.vs(s32,s8,s8) | 599.34 GOPS | | asimd_dp | dp4a.vv(s32,s8,s8) | 600.34 GOPS | | asimd_dp | dp4a.vs(u32,u8,u8) | 600.04 GOPS | | asimd_dp | dp4a.vv(u32,u8,u8) | 598.17 GOPS | | asimd_hp | fmla.vs(fp16,fp16,fp16) | 298.94 GFLOPS | | asimd_hp | fmla.vv(fp16,fp16,fp16) | 298.91 GFLOPS | | asimd | fmla.vs(f32,f32,f32) | 150 GFLOPS | | asimd | fmla.vv(f32,f32,f32) | 150.08 GFLOPS | | asimd | fmla.vs(f64,f64,f64) | 75.046 GFLOPS | | asimd | fmla.vv(f64,f64,f64) | 75.034 GFLOPS | ----------------------------------------------------------------