add x86_64 optimization to Makefile

[joneqdaniel]

• use -march=native if not cross-compiling
• use -mfpmath=sse if target is x86_64
• add -O3 -ftree-vectorize -fopenmp -fopenmp-simd to OPTIMIZE
• add -fdata-sections -ffunction-sections to RELEASE_CFLAGS

[ensiform]

What's the rationale behind adding openmp et all?

[ec-]

Please provide rationale and performance comparison

[joneqdaniel]

The reason for using OpenMP is platform independent SIMD instruction set and compiler support.
OpenMP is the only platform independent SIMD that is supported by MSVC compilers.

Compile the following arbitrary length SIMD dot product on any OpenMP supported target platform using any compiler supported.

• OpenMP Compilers & Tools

Compile with CFLAGS="-march=native -ftree-vectorize -fopenmp -fopenmp-simd -std=std2y" adding -O0 -to -O3 and view the assembler generated with -S for different target platforms and optimization levels"

#include <stdlib.h>
#include <stdint.h>
#include <stddef.h>
#include <stdbool.h>

#include "q_shared.h"

#ifdef __clang__
#define addr(a) (__typeof__((a)[0])*) __builtin_addressof((a))
#else
#define addr(a) &(a)[0]
#endif

/* dup - non-parallel memcpy supporting array and vector types */
#define dup(n,dst,src) \
({ \
    _Pragma("omp simd") \
    for(size_t i = 0; i < n; i++) \
        (dst)[i] = (src)[i]; \
})

/* dot - non-parallel scalar product supporting array and vector types */
#define dot(n,a,b,i,t) \
({ \
    t dst = (t)i; \
    _Pragma("omp simd reduction(+:dst)") \
    for(size_t j = 0; j < n; j++) \
        dst += (t)(a)[j] * (t)(b)[j]; \
    dst; \
})

int main(int argc, char** argv)
{
        evec(4,vec_t) a = { 0, 1, 2, 3 };
        vec_t b[3];
        dup(3,b,a);
        vec_t c = dot(3,a,b,0,vec_t);
        printf("%f\n", c);
        exit(EXIT_SUCCESS);
}

[ec-]

Did you tried compiling code from PR?
Not counting compilation errors, passing -march=native is simply not acceptable because it will make release binaries incompatible for A LOT of systems if compiled with AVX512 on a host for example.
Also I'd like to see performance measurements for an actual Quake3e code

[joneqdaniel]

passing -march=native is simply not acceptable because it will make release binaries incompatible for A LOT of systems if compiled with AVX512 on a host for example.

Adding target platform dependent SIMD restrictions to the RELEASE_CFLAGS is recommended, e.g. -mno-avx512

[ec-]

Adding target platform dependent SIMD restrictions to the RELEASE_CFLAGS is recommended, e.g. -mno-avx512

You're just changing ceiling which bottom is at SSE2 for x86_64, same for non-x86 architectures.

Please fix compilation errors in a first order, do not submit non-working code here

[KG7x]

ioquake/ioq3#869
ioquake/ioq3#870

Y can try compile use build in local version before merge

[joneqdaniel]

Y can try compile use build in local version before merge

only .gitignore is usable at these links.

[joneqdaniel]

workflow should complete now with #warning being removed

[ec-]

It breaks VC2005 build as well, not present in GitHub Actions though.

Also could you measure actual performance difference?

[joneqdaniel]

I've checked in game performance with Intel UHD Graphics 630 and NVIDIA GeForce RTX 2060/PCIe/SSE2 card using opengl1 renderer and map Q3DM1.

Intel/Nvidia framerate compare

• High: 23/140

seta r_fbo 1
seta r_vbo 1
seta r_bloom 1
seta r_hdr 1
seta r_ext_supersample 1
seta r_ext_multisample 8 

• Low: 500/200

seta r_fbo 0
seta r_vbo 0
seta r_bloom 0
seta r_hdr 0
seta r_ext_supersample 0
seta r_ext_multisample 0

• FBO/VBO only: 300/200

seta r_fbo 1
seta r_vbo 1
seta r_bloom 0
seta r_hdr 0
seta r_ext_supersample 0
seta r_ext_multisample 0

I'll get +20 with using the aligned one.
I still need to compare the vector type differences and write a profiling benchmark.

[joneqdaniel]

It breaks VC2005 build as well

Please submit and post a link to the error log.
Could you please trigger workflow run again. So I am able to check that MSVC compiles except VC2005.

[ec-]

Please test changes on your own GitHub Actions / compilers first, I will not approve test runs anymore

[ec-]

I'll get +20 with using the aligned one.

+20 from what base, what those High: 23/140 numbers means?
Did you measured difference that came with -O3 -mfpmath=sse + forced inlining alone?

[ensiform]

Fyi openmp support is excluded in 2005 and 2008 express editions, may need to be extra download with regular too.

[joneqdaniel]

Yes, Visual Studio 2005/2008 Express Editions need the extra download, while the Standard Editions include it.

[joneqdaniel]

Please approve another workflow run.
I'll look into coding a benchmark example soon.

[KG7x]

Please approve another workflow run. I'll look into coding a benchmark example soon.

You havent even launched actions in your GitHub repository once in all this time...
How y can сonfirm that you fixed this and there will be no warnings or errors?

[joneqdaniel]

Fixed mispelled MSVC syntax here!

[KG7x]

Wow now start build it in you repo and check it

[joneqdaniel]

No, there are no action workflow builds possible in a fork.

[joneqdaniel]

All the scripts are not setup.

[joneqdaniel]

Ah, I see the run workflow now.

[joneqdaniel]

thank you.

[ec-]

Sorry, that's enough