From 901aae99f21846953d817163dd418309dda21b3e Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Branimir=20Karad=C5=BEi=C4=87?= Date: Sun, 17 Jul 2016 11:06:48 -0700 Subject: [PATCH] Refactor SIMD langext. --- include/bx/float4_langext.h | 214 ++++++++++++++++++------------------ include/bx/float4_t.h | 40 +++++-- 2 files changed, 137 insertions(+), 117 deletions(-) diff --git a/include/bx/float4_langext.h b/include/bx/float4_langext.h index 2214df6..1e1c2ba 100644 --- a/include/bx/float4_langext.h +++ b/include/bx/float4_langext.h @@ -47,9 +47,9 @@ namespace bx #define ELEMw 3 #define IMPLEMENT_SWIZZLE(_x, _y, _z, _w) \ template<> \ - BX_FLOAT4_FORCE_INLINE float4_t float4_swiz_##_x##_y##_z##_w(float4_t _a) \ + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_swiz_##_x##_y##_z##_w(float4_langext_t _a) \ { \ - float4_t result; \ + float4_langext_t result; \ result.vf = __builtin_shufflevector(_a.vf, _a.vf, ELEM##_x, ELEM##_y, ELEM##_z, ELEM##_w); \ return result; \ } @@ -64,7 +64,7 @@ namespace bx #define IMPLEMENT_TEST(_xyzw, _mask) \ template<> \ - BX_FLOAT4_FORCE_INLINE bool float4_test_any_##_xyzw(float4_t _test) \ + BX_FLOAT4_FORCE_INLINE bool float4_test_any_##_xyzw(float4_langext_t _test) \ { \ uint32_t tmp = ( (_test.uxyzw[3]>>31)<<3) \ | ( (_test.uxyzw[2]>>31)<<2) \ @@ -75,7 +75,7 @@ namespace bx } \ \ template<> \ - BX_FLOAT4_FORCE_INLINE bool float4_test_all_##_xyzw(float4_t _test) \ + BX_FLOAT4_FORCE_INLINE bool float4_test_all_##_xyzw(float4_langext_t _test) \ { \ uint32_t tmp = ( (_test.uxyzw[3]>>31)<<3) \ | ( (_test.uxyzw[2]>>31)<<2) \ @@ -104,114 +104,114 @@ IMPLEMENT_TEST(xyzw , 0xf); #undef IMPLEMENT_TEST template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_xyAB(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_xyAB(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 0, 1, 4, 5); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_ABxy(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_ABxy(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 4, 5, 0, 1); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_CDzw(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_CDzw(float4_langext_t _a, float4_langext_t _b) { - float4_t result; - result.vf = __builtin_shufflevector(_a.vf, _b.vf, 5, 7, 2, 3); + float4_langext_t result; + result.vf = __builtin_shufflevector(_a.vf, _b.vf, 6, 7, 2, 3); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_zwCD(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_zwCD(float4_langext_t _a, float4_langext_t _b) { - float4_t result; - result.vf = __builtin_shufflevector(_a.vf, _b.vf, 2, 3, 5, 7); + float4_langext_t result; + result.vf = __builtin_shufflevector(_a.vf, _b.vf, 2, 3, 6, 7); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_xAyB(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_xAyB(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 0, 4, 1, 5); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_yBxA(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_yBxA(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 1, 5, 0, 4); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_zCwD(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_zCwD(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 2, 6, 3, 7); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_CzDw(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_CzDw(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 6, 2, 7, 3); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_xAzC(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_xAzC(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 0, 4, 2, 6); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_shuf_yBwD(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_shuf_yBwD(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = __builtin_shufflevector(_a.vf, _b.vf, 1, 5, 3, 7); return result; } template<> - BX_FLOAT4_FORCE_INLINE float float4_x(float4_t _a) + BX_FLOAT4_FORCE_INLINE float float4_x(float4_langext_t _a) { return _a.fxyzw[0]; } template<> - BX_FLOAT4_FORCE_INLINE float float4_y(float4_t _a) + BX_FLOAT4_FORCE_INLINE float float4_y(float4_langext_t _a) { return _a.fxyzw[1]; } template<> - BX_FLOAT4_FORCE_INLINE float float4_z(float4_t _a) + BX_FLOAT4_FORCE_INLINE float float4_z(float4_langext_t _a) { return _a.fxyzw[2]; } template<> - BX_FLOAT4_FORCE_INLINE float float4_w(float4_t _a) + BX_FLOAT4_FORCE_INLINE float float4_w(float4_langext_t _a) { return _a.fxyzw[3]; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_ld(const void* _ptr) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_ld(const void* _ptr) { const uint32_t* input = reinterpret_cast(_ptr); - float4_t result; + float4_langext_t result; result.uxyzw[0] = input[0]; result.uxyzw[1] = input[1]; result.uxyzw[2] = input[2]; @@ -220,7 +220,7 @@ IMPLEMENT_TEST(xyzw , 0xf); } template<> - BX_FLOAT4_FORCE_INLINE void float4_st(void* _ptr, float4_t _a) + BX_FLOAT4_FORCE_INLINE void float4_st(void* _ptr, float4_langext_t _a) { uint32_t* result = reinterpret_cast(_ptr); result[0] = _a.uxyzw[0]; @@ -230,14 +230,14 @@ IMPLEMENT_TEST(xyzw , 0xf); } template<> - BX_FLOAT4_FORCE_INLINE void float4_stx(void* _ptr, float4_t _a) + BX_FLOAT4_FORCE_INLINE void float4_stx(void* _ptr, float4_langext_t _a) { uint32_t* result = reinterpret_cast(_ptr); result[0] = _a.uxyzw[0]; } template<> - BX_FLOAT4_FORCE_INLINE void float4_stream(void* _ptr, float4_t _a) + BX_FLOAT4_FORCE_INLINE void float4_stream(void* _ptr, float4_langext_t _a) { uint32_t* result = reinterpret_cast(_ptr); result[0] = _a.uxyzw[0]; @@ -247,109 +247,109 @@ IMPLEMENT_TEST(xyzw , 0xf); } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_ld(float _x, float _y, float _z, float _w) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_ld(float _x, float _y, float _z, float _w) { - float4_t result; - result.vf = { _x, _y, _z, _w }; + float4_langext_t result; + result.vf = (float __attribute__((vector_size(16)))){ _x, _y, _z, _w }; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_ild(uint32_t _x, uint32_t _y, uint32_t _z, uint32_t _w) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_ild(uint32_t _x, uint32_t _y, uint32_t _z, uint32_t _w) { - float4_t result; - result.vu = { _x, _y, _z, _w }; + float4_langext_t result; + result.vu = (uint32_t __attribute__((vector_size(16)))){ _x, _y, _z, _w }; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_splat(const void* _ptr) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_splat(const void* _ptr) { const uint32_t val = *reinterpret_cast(_ptr); - float4_t result; - result.vu = { val, val, val, val }; + float4_langext_t result; + result.vu = (uint32_t __attribute__((vector_size(16)))){ val, val, val, val }; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_splat(float _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_splat(float _a) { - return float4_ld(_a, _a, _a, _a); + return float4_ld(_a, _a, _a, _a); } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_isplat(uint32_t _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_isplat(uint32_t _a) { - return float4_ild(_a, _a, _a, _a); + return float4_ild(_a, _a, _a, _a); } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_zero() + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_zero() { - return float4_ild(0, 0, 0, 0); + return float4_ild(0, 0, 0, 0); } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_itof(float4_t _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_itof(float4_langext_t _a) { - float4_t result; + float4_langext_t result; result.vf = __builtin_convertvector(_a.vi, float __attribute__((vector_size(16))) ); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_ftoi(float4_t _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_ftoi(float4_langext_t _a) { - float4_t result; + float4_langext_t result; result.vi = __builtin_convertvector(_a.vf, int32_t __attribute__((vector_size(16))) ); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_round(float4_t _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_round(float4_langext_t _a) { - const float4_t tmp = float4_ftoi(_a); - const float4_t result = float4_itof(tmp); + const float4_langext_t tmp = float4_ftoi(_a); + const float4_langext_t result = float4_itof(tmp); return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_add(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_add(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = _a.vf + _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_sub(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_sub(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = _a.vf - _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_mul(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_mul(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = _a.vf * _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_div(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_div(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vf = _a.vf / _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_sqrt(float4_t _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_sqrt(float4_langext_t _a) { - float4_t result; + float4_langext_t result; result.vf[0] = sqrtf(_a.vf[0]); result.vf[1] = sqrtf(_a.vf[1]); result.vf[2] = sqrtf(_a.vf[2]); @@ -358,9 +358,9 @@ IMPLEMENT_TEST(xyzw , 0xf); } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_rsqrt_est(float4_t _a) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_rsqrt_est(float4_langext_t _a) { - float4_t result; + float4_langext_t result; result.vf[0] = 1.0f / sqrtf(_a.vf[0]); result.vf[1] = 1.0f / sqrtf(_a.vf[1]); result.vf[2] = 1.0f / sqrtf(_a.vf[2]); @@ -369,144 +369,146 @@ IMPLEMENT_TEST(xyzw , 0xf); } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_cmpeq(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_cmpeq(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vf == _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_cmplt(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_cmplt(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vf < _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_cmple(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_cmple(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vf <= _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_cmpgt(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_cmpgt(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vf > _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_cmpge(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_cmpge(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vf >= _b.vf; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_and(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_and(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vu = _a.vu & _b.vu; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_andc(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_andc(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vu = _a.vu & ~_b.vu; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_or(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_or(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vu = _a.vu | _b.vu; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_xor(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_xor(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vu = _a.vu ^ _b.vu; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_sll(float4_t _a, int _count) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_sll(float4_langext_t _a, int _count) { - float4_t result; - const float4_t count = float4_isplat(_count); + float4_langext_t result; + const float4_langext_t count = float4_isplat(_count); result.vu = _a.vu << count.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_srl(float4_t _a, int _count) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_srl(float4_langext_t _a, int _count) { - float4_t result; - const float4_t count = float4_isplat(_count); + float4_langext_t result; + const float4_langext_t count = float4_isplat(_count); result.vu = _a.vu >> count.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_sra(float4_t _a, int _count) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_sra(float4_langext_t _a, int _count) { - float4_t result; - const float4_t count = float4_isplat(_count); + float4_langext_t result; + const float4_langext_t count = float4_isplat(_count); result.vi = _a.vi >> count.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_icmpeq(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_icmpeq(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vi == _b.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_icmplt(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_icmplt(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vi < _b.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_icmpgt(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_icmpgt(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vi > _b.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_iadd(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_iadd(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vi + _b.vi; return result; } template<> - BX_FLOAT4_FORCE_INLINE float4_t float4_isub(float4_t _a, float4_t _b) + BX_FLOAT4_FORCE_INLINE float4_langext_t float4_isub(float4_langext_t _a, float4_langext_t _b) { - float4_t result; + float4_langext_t result; result.vi = _a.vi - _b.vi; return result; } + typedef float4_langext_t float4_t; + } // namespace bx #endif // BX_FLOAT4_LANGEXT_H_HEADER_GUARD diff --git a/include/bx/float4_t.h b/include/bx/float4_t.h index 70bad40..a449546 100644 --- a/include/bx/float4_t.h +++ b/include/bx/float4_t.h @@ -11,12 +11,19 @@ #define BX_FLOAT4_FORCE_INLINE BX_FORCE_INLINE #define BX_FLOAT4_INLINE inline +#define BX_FLOAT4_SSE 0 +#define BX_FLOAT4_AVX 0 +#define BX_FLOAT4_NEON 0 +#define BX_FLOAT4_LANGEXT 0 + #if defined(__SSE2__) || (BX_COMPILER_MSVC && (BX_ARCH_64BIT || _M_IX86_FP >= 2) ) # include // __m128i # if defined(__SSE4_1__) # include # endif // defined(__SSE4_1__) # include // __m128 +# undef BX_FLOAT4_SSE +# define BX_FLOAT4_SSE 1 namespace bx { @@ -26,6 +33,8 @@ namespace bx #elif defined(__ARM_NEON__) && !BX_COMPILER_CLANG # include +# undef BX_FLOAT4_NEON +# define BX_FLOAT4_NEON 1 namespace bx { @@ -38,10 +47,12 @@ namespace bx && !BX_PLATFORM_IOS \ && BX_CLANG_HAS_EXTENSION(attribute_ext_vector_type) # include +# undef BX_FLOAT4_LANGEXT +# define BX_FLOAT4_LANGEXT 1 namespace bx { - typedef union float4_langext_t + union float4_langext_t { float __attribute__((vector_size(16))) vf; int32_t __attribute__((vector_size(16))) vi; @@ -50,19 +61,19 @@ namespace bx int32_t ixyzw[4]; uint32_t uxyzw[4]; - } float4_langext_t; + }; } // namespace bx #endif // namespace bx { - typedef union float4_ref_t + union float4_ref_t { float fxyzw[4]; int32_t ixyzw[4]; uint32_t uxyzw[4]; - } float4_ref_t; + }; } // namespace bx namespace bx @@ -351,16 +362,23 @@ IMPLEMENT_TEST(xyzw); } // namespace bx -#if defined(__SSE2__) || (BX_COMPILER_MSVC && (BX_ARCH_64BIT || _M_IX86_FP >= 2) ) +#if BX_FLOAT4_SSE # include "float4_sse.h" -#elif defined(__ARM_NEON__) && !BX_COMPILER_CLANG +#endif // BX_FLOAT4_SSE + +#if BX_FLOAT4_NEON # include "float4_neon.h" -#elif BX_COMPILER_CLANG \ - && !BX_PLATFORM_EMSCRIPTEN \ - && !BX_PLATFORM_IOS \ - && BX_CLANG_HAS_EXTENSION(attribute_ext_vector_type) +#endif // BX_FLOAT4_NEON + +#if BX_FLOAT4_LANGEXT # include "float4_langext.h" -#else +#endif // BX_FLOAT4_LANGEXT + +#if !( BX_FLOAT4_SSE \ + || BX_FLOAT4_AVX \ + || BX_FLOAT4_NEON \ + || BX_FLOAT4_LANGEXT \ + ) # ifndef BX_FLOAT4_WARN_REFERENCE_IMPL # define BX_FLOAT4_WARN_REFERENCE_IMPL 0 # endif // BX_FLOAT4_WARN_REFERENCE_IMPL