Remove excessive floating-point divides

[heshpdx]

Loft the loop-invariant divide outside the hot loops, and/or invert the variable to turn FDIV into FMUL.

Most CPUs are slower at FP division compared to FP multiplication. This should provide some uplift in performance. I was testing with the integer models.

[stweil]

Do you have timing values from your tests?

[heshpdx]

Do you have timing values from your tests?

It will be CPU specific, but I see +10% on my Ampere Altra.

[stweil]

It will be CPU specific, but I see +10% on my Ampere Altra.

That's a very significant improvement! I wonder how this ARM64 cpu compares to Intel / AMD cpus with Tesseract recognition and training.

[heshpdx]

If there are standard tests that you run, please do share the results.

I was using -l deu --tessdata-dir ./tessdata_orig --oem 0 and -l Arabic --tessdata-dir ./tessdata_fast. The -l rus --tessdata-dir ./tessdata_orig --oem 2 did not show much improvement.

[stweil]

Does Ampere Altra offer additional opcodes which could be used to make Tesseract's neural network code faster? We currently use Neon code for ARM64 (see src/arch/*neon.cpp).

[stweil]

If there are standard tests that you run, please do share the results.

You can run make check (after installing required packages, repositories and git submodules) and compare the times for the single tests. lstm_test is the test with the longest execution time.

Here are my results on a Mac mini M2 for running time ./lstm_test:

# git main branch, extract from lstm_test.log and log message from `time`.
[==========] 11 tests from 1 test suite ran. (278833 ms total)
./lstm_test  274,78s user 2,87s system 99% cpu 4:38,88 total

# git main branch with PR applied, extract from lstm_test.log and log message from `time`.
[==========] 11 tests from 1 test suite ran. (276981 ms total)
./lstm_test  273,60s user 2,50s system 99% cpu 4:37,03 total