src: make coefficients constant

[lyakh]

Make coefficients and their tables constant.

[singalsu]

Reminder to myself, update the generator script to produce header files with this change.

[lgirdwood]

@singalsu btw, all the SRC coefficients are loaded into uncache memory IIUC i.e. from binutils nm

400a9ca8 D src_in_fs
400a9cc4 D src_int16_0_0_0_0
400a9cec D src_int16_1_1_0_0
400a9f1c D src_int16_1_2_1814_5000
400a9ef4 D src_int16_1_3_1814_5000
400a9ecc D src_int16_1_6_1814_5000
400a9d8c D src_int16_20_21_1667_5000
400a9ea4 D src_int16_2_1_1814_5000
400a9d64 D src_int16_21_20_1667_5000
400a9e7c D src_int16_2_3_1814_5000
400a9d3c D src_int16_24_25_1814_5000
400a9d14 D src_int16_25_24_1814_5000
400a9e54 D src_int16_3_1_1814_5000
400a9e2c D src_int16_3_2_1814_5000
400a9e04 D src_int16_6_1_1814_5000
400a9ddc D src_int16_7_8_1814_5000
400a9db4 D src_int16_8_7_1814_5000
400aa098 D src_lite_tr
400a9c8c D src_out_fs
400a9bc8 D src_table1
400a9b04 D src_table2
400aa090 D src_tr

This means we need to ucache_to_cache then when we set up the pointers for coeffs.

[lyakh]

@singalsu btw, all the SRC coefficients are loaded into uncache memory IIUC i.e. from binutils nm

400a9ca8 D src_in_fs
400a9cc4 D src_int16_0_0_0_0
400a9cec D src_int16_1_1_0_0
400a9f1c D src_int16_1_2_1814_5000
400a9ef4 D src_int16_1_3_1814_5000
400a9ecc D src_int16_1_6_1814_5000
400a9d8c D src_int16_20_21_1667_5000
400a9ea4 D src_int16_2_1_1814_5000
400a9d64 D src_int16_21_20_1667_5000
400a9e7c D src_int16_2_3_1814_5000
400a9d3c D src_int16_24_25_1814_5000
400a9d14 D src_int16_25_24_1814_5000
400a9e54 D src_int16_3_1_1814_5000
400a9e2c D src_int16_3_2_1814_5000
400a9e04 D src_int16_6_1_1814_5000
400a9ddc D src_int16_7_8_1814_5000
400a9db4 D src_int16_8_7_1814_5000
400aa098 D src_lite_tr
400a9c8c D src_out_fs
400a9bc8 D src_table1
400a9b04 D src_table2
400aa090 D src_tr

This means we need to ucache_to_cache then when we set up the pointers for coeffs.

@lgirdwood this is before this commit, right? After it I see these in .rodata with cached aliases

[singalsu]

@lgirdwood @lyakh Wait, do you mean that const C arrays are potentially corrupt if accessed without those address conversions?

[singalsu]

Also, fast coefficients access is key for good performance. If memory where these are located is slow, the tables that the active x to y kHz rate conversion uses should be copied to a memory where the SIMD MAC instructions work efficiently. Because the tables are very large like 100 - 200 kB it would make sense to keep them in a large slow memory and copy to fastest RAM only the much smaller that time needed set.

[lyakh]

@lgirdwood @lyakh Wait, do you mean that const C arrays are potentially corrupt if accessed without those address conversions?

so weird, I thought I'd replied, sorry. No, it isn't about corruption. And I don't think any conversion is needed. After this commit I see all those arrays accessed via cached aliases, but maybe I'm missing some build configuration @lgirdwood ?

BTW, there's one more thing to check - explicit type conversions like

sof/src/audio/src/src_hifi2ep.c

Line 59 in 45dc968

coefp = (ae_p16x2s *)cp;

- that's already a bug. That one explicitly removes the const qualifier from exactly those coefficients IIUC, so it can then in principle be used to corrupt them, yes. I assume we have a few of such conversions in the sources...

[lyakh]

Replying to myself: I found -Wcast-qual - @singalsu please try building the sources with it, and maybe @marc-hb we should add it to SOF in general

[lyakh]

CI: a known suspend-resume problem in https://sof-ci.01.org/sofpr/PR8829/build2554/devicetest/index.html and a supposedly unrelated fuzzer failure. "main" Zephyr build failures are known too

[lgirdwood]

This means we need to ucache_to_cache then when we set up the pointers for coeffs.

@lgirdwood this is before this commit, right? After it I see these in .rodata with cached aliases

ok, great good to know RO is properly mapped here.

[singalsu]

Replying to myself: I found -Wcast-qual - @singalsu please try building the sources with it, and maybe @marc-hb we should add it to SOF in general

I tried but I have no idea how to fix container_of() macro.

[lyakh]

Replying to myself: I found -Wcast-qual - @singalsu please try building the sources with it, and maybe @marc-hb we should add it to SOF in general

I tried but I have no idea how to fix container_of() macro.

@singalsu we could add a

#define const_container_of(ptr, type, member) \
	({const typeof(((type *)0)->member)*__memberptr = (ptr); \
	(const type *)((const char *)__memberptr - offsetof(type, member)); })

and use it for const pointers instead of the usual container_of()

[singalsu]

Thanks @lyakh I'll try it if I can proceed with it!