This is a modification of the Tamp compression library created by Brian Pugh, aggressively speed-optimized for use as a C/C++ library on ESP32 MCUs.
The C implementation of the original Tamp library is made into a component for use in ESP-IDF builds and speed-optimized for 32-bit architectures like the ESP32s', including faster pattern-match search functions.
While the original Tamp C code is compatible with most (/all?) architectures, this optimized variant leverages the ESP32s' ability to do unaligned 16- and 32-bit memory accesses. It should be able to run on architectures other than ESP32s (e.g. other RISC-V CPUs), but it will not work on chips which don't support unaligned memory accesses (e.g. the Raspberry Pi Pico RP2040).
The performance optimizations work out to about 2.5-4x faster compression on ESP32s (both Xtensa and RISC-V, i.e. ESP32, ESP32-C2, ESP32-S3, &c.). When built for an ESP32-S3, the MCU's SIMD instructions ("PIE") are used which increases compression speed/ throughput by a factor of 8-20x.
When building the C library as a component in an ESP-IDF build, you can choose to build the original or the ESP32-optimized variant in menuconfig.
Tamp can deliver decent compression with minimal memory and CPU requirements.
Input: 5614 bytes of text
Compressed output in bytes vs. window size ("window bits (bytes)"), i.e. RAM requirement
| Uncompressed | 8 (256b) | 9 (512b) | 10 (1kb) | 11 (2kb) | 12 (4kb) | |
|---|---|---|---|---|---|---|
| Bytes | 5614 | 3198 | 2801 | 2650 | 2581 | 2571 |
| % | 100,0 % | 56,9 % | 49,9 % | 47,2 % | 46,0 % | 45,8 % |
| Compression | 1,0 x | 1,8 x | 2,0 x | 2,1 x | 2,2 x | 2,2 x |
Input: Espressif logo in 24-bit BMP format, 25862 bytes
Compressed output in bytes vs. window size ("window bits (bytes)")
| Uncompressed | 8 (256b) | 9 (512b) | 10 (1kb) | 11 (2kb) | 12 (4kb) | 13 (8kb) | |
|---|---|---|---|---|---|---|---|
| Bytes | 25862 | 7720 | 7772 | 6539 | 6517 | 6601 | 6590 |
| % | 100,0 % | 29,9 % | 30,1 % | 25,3 % | 25,2 % | 25,5 % | 25,5 % |
| Compression | 1,0 x | 3,4 x | 3,3 x | 4,0 x | 4,0 x | 3,9 x | 3,9 x |
As to the overall compression efficiency, 7zip in "Ultra" mode compresses the same file down to 3022 bytes (11,7%), which should be a rough approximation of the entropy of that file, i.e. the theoretical limit of what is possible.
Performance on an ESP32-S3 (Xtensa) @ 240 MHz:
Input: Espressif logo in 24-bit BMP format, 25862 bytes
Original C version:
| 8 | 9 | 10 | 11 | 12 | 13 | |
|---|---|---|---|---|---|---|
| CPU time | 103,5 ms | 182,1 ms | 252,7 ms | 420,7 ms | 782,1 ms | 1419,1 ms |
| Throughput | 244,0 kb/s | 138,7 kb/s | 99,9 kb/s | 60,0 kb/s | 32,3 kb/s | 17,8 kb/s |
Optimized C/C++ variant:
| 8 | 9 | 10 | 11 | 12 | 13 | |
|---|---|---|---|---|---|---|
| CPU time | 27,0 ms | 45,1 ms | 60,2 ms | 102,6 ms | 190,4 ms | 344,4 ms |
| Throughput | 936,5 kb/s | 560,5 kb/s | 419,7 kb/s | 246,2 kb/s | 132,6 kb/s | 73,3 kb/s |
| Speed (rel) | 3,8 x | 4,0 x | 4,2 x | 4,1 x | 4,1 x | 4,1 x |
ESP32-S3 SIMD/PIE variant:
| 8 | 9 | 10 | 11 | 12 | 13 | |
|---|---|---|---|---|---|---|
| CPU time | 9,7 ms | 11,9 ms | 13,2 ms | 19,3 ms | 30,4 ms | 51,4 ms |
| Throughput | 2598,3 kb/s | 2126,4 kb/s | 1908,6 kb/s | 1310,2 kb/s | 831,8 kb/s | 491,4 kb/s |
| Speed (rel) | 10,6 x | 15,3 x | 19,1 x | 21,8 x | 25,8 x | 27,6 x |
An ESP32-C3 (RISC-V, max. 160 MHz) consistently runs the original variant about 5% faster and the optimized C/C++ variant about 10% faster per CPU clock cycle than the ESP32-S3; so the C3 is, for this workload, a little more efficient than the S3.