Unroll some tight loops

[howard0su]

No description provided.

[fventuri]

Good idea Howard!
I don't know how C++ compilers work these days, but perhaps a block like this:

            *output++ = float(*input++);
            *output++ = float(*input++);
            *output++ = float(*input++);
            *output++ = float(*input++);

could be replaced by something like this:

            const int16_t *in = input + 4 * m;
            float *out = output + 4 * m;
            out[0] = float(int[0]);
            out[1] = float(int[1]);
            out[2] = float(int[2]);
            out[3] = float(int[3]);

I don't know if from the point of the compiler the second option could use SIMD instructions, since there are no output+_ and input++ around, but perhaps the compiler figures it out since it is idiomatic.

There's also the volk library (https://www.libvolk.org/) that has a couple of functions called 'volk_16i_s32f_convert_32f' (https://www.libvolk.org/doxygen/volk_16i_s32f_convert_32f.html) and 'volk_16ic_convert_32fc' (https://www.libvolk.org/doxygen/volk_16ic_convert_32fc.html), which are optimized for different types of hardware, but I am not sure if its license (GPL v3) is compatible with this project's license.

Franco

[howard0su]

I got the idea from CMSIS-DSP. Ideally, hand write instruction can do a better job but hard to adopt to different SIMD solutions. I will do more experiment to see how difference they are.

…

On Sat, Jul 6, 2024 at 10:54 PM Franco Venturi ***@***.***> wrote: Good idea Howard! I don't know how C++ compilers work these days, but perhaps a block like this: *output++ = float(*input++); *output++ = float(*input++); *output++ = float(*input++); *output++ = float(*input++); could be replaced by something like this: const int16_t *in = input + 4 * m; float *out = output + 4 * m; out[0] = float(int[0]); out[1] = float(int[1]); out[2] = float(int[2]); out[3] = float(int[3]); I don't know if from the point of the compiler the second option could use SIMD instructions, since there are no output+_ and input++ around, but perhaps the compiler figures it out since it is idiomatic. There's also the volk library (https://www.libvolk.org/) that has a couple of functions called 'volk_16i_s32f_convert_32f' ( https://www.libvolk.org/doxygen/volk_16i_s32f_convert_32f.html) and 'volk_16ic_convert_32fc' ( https://www.libvolk.org/doxygen/volk_16ic_convert_32fc.html), which are optimized for different types of hardware, but I am not sure if its license (GPL v3) is compatible with this project's license. Franco — Reply to this email directly, view it on GitHub <#235 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAF3GRG3H7J2CTQAXZGSAVLZLAAL7AVCNFSM6AAAAABKOOA3BSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEMJRG44DSMBXHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

-- -Howard

[cozycactus]

what do you think about fast int16_t to float conversion like this https://github.com/m-ou-se/floatconv
when i did perf it showed conversion takes many cpu %

[cozycactus]

https://blog.m-ou.se/floats/

[howard0su]

This is interesting blog. I will look into it and port i16_to_float over.

…

On Sun, Jul 7, 2024 at 5:22 AM Ruslan Migirov ***@***.***> wrote: https://blog.m-ou.se/floats/ — Reply to this email directly, view it on GitHub <#235 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAF3GRDEZWWXJHMFTVMP6RLZLBNXRAVCNFSM6AAAAABKOOA3BSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEMJRHE3TEOBRGM> . You are receiving this because you authored the thread.Message ID: ***@***.***>

-- -Howard

[cozycactus]

i tried this optimization and i have found only convert_float one showed under one percent better cpu usage on r2iq thread...