WIP: Wasm simd128 support

[myfreeer]

See #282 for previous discussion.

TODOs

• Inline or rewrite sse2 wrappers borrowed from emscripten

  • _mm_cvtss_f32
  • _mm_movehl_ps
  • _mm_movelh_ps
  • _mm_sqrt_ss
  • _mm_rcp_ps
  • _mm_storeu_ps
  • _mm_store_ss
  • _MM_TRANSPOSE4_PS

• Guard againist conditions when both sse2 and simd128 enabled by Emscripten (WIP: Wasm simd128 support #286 (comment))

  • affine-mat.h
  • intrin.h
  • mat2.h
  • mat3.h
  • mat4.h
  • quat.h
  • vec4-ext.h
  • vec4.h

• Change in compile scripts for cross-compiling or a build guide

  • cmake
  • autoconf delayed since it could take too lang
  • meson delayed since it could take too lang

• CI-driven tests, cross compile all tests with or without simd128, using appveyor or Github Actions

  • target: wasi-sdk
  • target: wasi-sdk with simd128
  • target: emsdk
  • target: emsdk with simd128
  • target: emsdk with sse2 and simd128
  • runtime: wasmtime
  • runtime: wasmer

• Code style changes

Compiling

Currently the tests can be cross-compiled to wasi using clang and wasi-sdk with cmake arguments below:

-DCMAKE_C_FLAGS="-msimd128"
-DCMAKE_TOOLCHAIN_FILE=/path/to/wasi-sdk-19.0/share/cmake/wasi-sdk.cmake
-DWASI_SDK_PREFIX=/path/to/wasi-sdk-19.0
-DCGLM_USE_TEST=ON

Where /path/to/wasi-sdk-19.0/ is the path to extracted wasi sdk.

The cmake-wasm.yml shows an example for compiling and running tests on github actions.

[kzhsw]

The _mm_storeu_ps here is defined mostly like the build-in wasm_v128_store, and there two would compile to the same result in clang.

static __inline__ void __attribute__((__always_inline__, __nodebug__))
_mm_storeu_ps(float *__p, glmm_128 __a)
{
  struct __unaligned {
    glmm_128 __v;
  } __attribute__((__packed__, __may_alias__));
  ((struct __unaligned *)__p)->__v = __a;
}

static __inline__ void __DEFAULT_FN_ATTRS wasm_v128_store(void *__mem,
                                                          v128_t __a) {
  // UB-free unaligned access copied from xmmintrin.h
  struct __wasm_v128_store_struct {
    __v128_u __v;
  } __attribute__((__packed__, __may_alias__));
  ((struct __wasm_v128_store_struct *)__mem)->__v = __a;
}

[gottfriedleibniz]

Another potential edge-case that may require handling is compiling w/ Emscripten, e.g., emcc -msimd128 -msse -msse2 ... which will enable both CGLM_SIMD_x86 and CGLM_SIMD_WASM.

[myfreeer]

Another potential edge-case that may require handling is compiling w/ Emscripten, e.g., emcc -msimd128 -msse -msse2 ... which will enable both CGLM_SIMD_x86 and CGLM_SIMD_WASM.

This could be guarded by some extra check in macros like this, a lot more similar things should be changed.

diff --git a/include/cglm/simd/intrin.h b/include/cglm/simd/intrin.h 
index 17998f5..bf8d119 100644                                        
--- a/include/cglm/simd/intrin.h                                     
+++ b/include/cglm/simd/intrin.h                                     
@@ -113,7 +113,7 @@                                                  
 #  endif
 #endif

-#if defined(CGLM_SIMD_x86) 
+#if defined(CGLM_SIMD_x86) && !defined(CGLM_SIMD_WASM)
 #  include "x86.h"
 #endif

diff --git a/include/cglm/vec4.h b/include/cglm/vec4.h
index d6ee080..e90a3c3 100644
--- a/include/cglm/vec4.h
+++ b/include/cglm/vec4.h 
@@ -181,10 +181,10 @@ glm_vec4_ucopy(vec4 v, vec4 dest) {
 CGLM_INLINE
 void
 glm_vec4_zero(vec4 v) {
-#if defined( __SSE__ ) || defined( __SSE2__ )
-  glmm_store(v, _mm_setzero_ps());
-#elif defined(__wasm__) && defined(__wasm_simd128__)
+#if defined(__wasm__) && defined(__wasm_simd128__)
   glmm_store(v, wasm_f32x4_const_splat(0.f));
+#elif defined( __SSE__ ) || defined( __SSE2__ )
+  glmm_store(v, _mm_setzero_ps());
 #elif defined(CGLM_NEON_FP)
   vst1q_f32(v, vdupq_n_f32(0.0f));
 #else

[recp]

@myfreeer great work! any new progress?

[myfreeer]

@myfreeer great work! any new progress?

Yes, github actions added for tests.

[recp]

@myfreeer awesome 🚀

[recp]

@myfreeer any new progress ? It would be nice to merge this asap

[myfreeer]

@myfreeer any new progress ? It would be nice to merge this asap

@recp
No, I was planning for support of autoconf and meson support, autotools failed to build tests, and meson needs some cross-file with absolute path, some extra helper scripts might be needed for generating this, meson would fail to build tests with maybe the wasm-ld issue. Maybe we should delay supporting those (maybe in another pr), and mark this as ready-for-review.
BTW, which tool is suitable for code style checking?

[recp]

No, I was planning for support of autoconf and meson support, autotools failed to build tests, and meson needs some cross-file with absolute path, some extra helper scripts might be needed for generating this, meson would fail to build tests with maybe the mesonbuild/meson#10187 issue. Maybe we should delay supporting those (maybe in another pr), and mark this as ready-for-review.

@myfreeer yes this is good approach I think. This PR became huge update, another PR[s] would be awesome, I hope we won't forget it :)

BTW, which tool is suitable for code style checking?

🤷‍♂️ if it will take long time, I or you could do it later

[recp]

@myfreeer the PR is merged 🚀 many thanks. Feel free to create PRs for additional improvements

[myfreeer]

Great! Also, it seems that the github actions ci needs to be manually enabled in this repo to make it work.

[recp]

I have just enabled the actions 👍