Benchmarks 2024 02 11 TFLM GCC

Setup

Simulator

• Spike (riscv-isa-sim ) (ISS, CPI=1)

Toolchains

• RISC-V GCC:
  • Scalar: TODO: version & url
  • Vector: TODO: version & url
  • Packed: Self compiled using patches found in https://github.com/riscv-collab/riscv-gcc/pull/258 and https://github.com/riscvarchive/riscv-binutils-gdb/pull/257

Models

• MLPerfTiny Benchmark

• TODO: others!

Package Versions

• MLonMCU : main

• TFLM : a549448bb234cf3fed15ad5dabf83d06f82326ce

• Spike : 0bc176b3fca43560b9e8586cdbc41cfde073e17a

• Spike PK : 7e9b671c0415dfd7b562ac934feb9380075d4aa2

Miscellaneous

• Used -Os flag for compilation.
• Benchmarks generated using MLonMCU deployment tool with minimal efforts.
• Memory metrics are reported in Bytes

Results (Framework: tflm, Backend: tflmi, Toolchain: gcc)

Audio Wake Words (aww)

Cycles (Speedup)	Total ROM (rel.)	Total RAM (rel.)	VLEN	Kernels	Mode	Arch	Auto-Vectorization
174698646 ( 0.1x )	132407 ( 0.878 )	36204 ( 1.0 )	0	TFLM	Reference	RV32GC	-
174698646 ( 0.1x )	132413 ( 0.878 )	36204 ( 1.0 )	128	TFLM	Reference	RV32GCV	Loop+SLP
174698646 ( 0.1x )	132413 ( 0.878 )	36204 ( 1.0 )	1024	TFLM	Reference	RV32GCV	Loop+SLP
157549999 ( 0.1x )	144774 ( 0.96 )	36148 ( 0.998 )	0	TFLM	Reference	RV32GCP	-
16644695 ( Base )	150798 ( Base )	36212 ( Base )	0	muRISCV-NN	Scalar	RV32GC	-
16644695 ( 1.0x )	150804 ( 1.0 )	36212 ( 1.0 )	128	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
16644695 ( 1.0x )	150804 ( 1.0 )	36212 ( 1.0 )	1024	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
7002129 ( 2.4x )	151322 ( 1.003 )	36212 ( 1.0 )	128	muRISCV-NN	Vector	RV32GCV	-
2571441 ( 6.5x )	151322 ( 1.003 )	36212 ( 1.0 )	1024	muRISCV-NN	Vector	RV32GCV	-
13498251 ( 1.2x )	162366 ( 1.077 )	36156 ( 0.998 )	0	muRISCV-NN	Scalar	RV32GCP	-
15939667 ( 1.0x )	164764 ( 1.093 )	36156 ( 0.998 )	0	muRISCV-NN	Packed	RV32GCP	-

Notes

• TFLM Reference kernels perform extremely bad with GCC (LLVM is more than 2-4x faster here)
• AutoVectorization disabled in GCC? (Check MLonMCU config!)
• Packed muRISC-V kernels have little speedup (1-1.2x) for CNNs (Even Scalar + RVP GCC performs better..., DNNs are fine)

Image Classification (resnet)

Cycles (Speedup)	Total ROM (rel.)	Total RAM (rel.)	VLEN	Kernels	Mode	Arch	Auto-Vectorization
745801113 ( 0.1x )	172997 ( 0.934 )	68968 ( 1.0 )	0	TFLM	Reference	RV32GC	-
745801113 ( 0.1x )	173013 ( 0.934 )	68968 ( 1.0 )	128	TFLM	Reference	RV32GCV	Loop+SLP
745801113 ( 0.1x )	173013 ( 0.934 )	68968 ( 1.0 )	1024	TFLM	Reference	RV32GCV	Loop+SLP
697912970 ( 0.1x )	185266 ( 1.0 )	68912 ( 0.999 )	0	TFLM	Reference	RV32GCP	-
80995160 ( Base )	185298 ( Base )	68960 ( Base )	0	muRISCV-NN	Scalar	RV32GC	-
80995160 ( 1.0x )	185330 ( 1.0 )	68960 ( 1.0 )	128	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
80995160 ( 1.0x )	185330 ( 1.0 )	68960 ( 1.0 )	1024	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
29960342 ( 2.7x )	186654 ( 1.007 )	68960 ( 1.0 )	128	muRISCV-NN	Vector	RV32GCV	-
8347502 ( 9.7x )	186654 ( 1.007 )	68960 ( 1.0 )	1024	muRISCV-NN	Vector	RV32GCV	-
62976509 ( 1.3x )	196964 ( 1.063 )	68904 ( 0.999 )	0	muRISCV-NN	Scalar	RV32GCP	-
68431268 ( 1.2x )	199958 ( 1.079 )	68904 ( 0.999 )	0	muRISCV-NN	Packed	RV32GCP	-

Anomaly Detection (toycar)

Cycles (Speedup)	Total ROM (rel.)	Total RAM (rel.)	VLEN	Kernels	Mode	Arch	Auto-Vectorization
3094956 ( 0.6x )	333908 ( 0.989 )	19432 ( 1.0 )	0	TFLM	Reference	RV32GC	-
3094956 ( 0.6x )	333914 ( 0.989 )	19432 ( 1.0 )	128	TFLM	Reference	RV32GCV	Loop+SLP
3094956 ( 0.6x )	333914 ( 0.989 )	19432 ( 1.0 )	1024	TFLM	Reference	RV32GCV	Loop+SLP
3097895 ( 0.6x )	346264 ( 1.026 )	19380 ( 0.997 )	0	TFLM	Reference	RV32GCP	-
1969300 ( Base )	337546 ( Base )	19432 ( Base )	0	muRISCV-NN	Scalar	RV32GC	-
1969300 ( 1.0x )	337552 ( 1.0 )	19432 ( 1.0 )	128	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
1969300 ( 1.0x )	337552 ( 1.0 )	19432 ( 1.0 )	1024	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
608151 ( 3.2x )	337800 ( 1.001 )	19432 ( 1.0 )	128	muRISCV-NN	Vector	RV32GCV	-
428279 ( 4.6x )	337800 ( 1.001 )	19432 ( 1.0 )	1024	muRISCV-NN	Vector	RV32GCV	-
1873278 ( 1.1x )	349708 ( 1.036 )	19380 ( 0.997 )	0	muRISCV-NN	Scalar	RV32GCP	-
942627 ( 2.1x )	351426 ( 1.041 )	19380 ( 0.997 )	0	muRISCV-NN	Packed	RV32GCP	-

Visual Wake Words (vww)

Cycles (Speedup)	Total ROM (rel.)	Total RAM (rel.)	VLEN	Kernels	Mode	Arch	Auto-Vectorization
495266208 ( 0.1x )	406111 ( 0.957 )	134520 ( 1.0 )	0	TFLM	Reference	RV32GC	-
495266207 ( 0.1x )	406117 ( 0.957 )	134520 ( 1.0 )	128	TFLM	Reference	RV32GCV	Loop+SLP
495266207 ( 0.1x )	406117 ( 0.957 )	134520 ( 1.0 )	1024	TFLM	Reference	RV32GCV	Loop+SLP
445892345 ( 0.1x )	418478 ( 0.986 )	134464 ( 1.0 )	0	TFLM	Reference	RV32GCP	-
49676633 ( Base )	424502 ( Base )	134528 ( Base )	0	muRISCV-NN	Scalar	RV32GC	-
49676633 ( 1.0x )	424508 ( 1.0 )	134528 ( 1.0 )	128	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
49676633 ( 1.0x )	424508 ( 1.0 )	134528 ( 1.0 )	1024	muRISCV-NN	Scalar	RV32GCV	Loop+SLP
21932927 ( 2.3x )	425026 ( 1.001 )	134528 ( 1.0 )	128	muRISCV-NN	Vector	RV32GCV	-
10503478 ( 4.7x )	425026 ( 1.001 )	134528 ( 1.0 )	1024	muRISCV-NN	Vector	RV32GCV	-
40750242 ( 1.2x )	436070 ( 1.027 )	134472 ( 1.0 )	0	muRISCV-NN	Scalar	RV32GCP	-
49184061 ( 1.0x )	438468 ( 1.033 )	134472 ( 1.0 )	0	muRISCV-NN	Packed	RV32GCP	-

Original data

Click here to download the raw files for this benchmark.