diff --git a/examples/intrinsics/knc.h b/examples/intrinsics/knc.h index e55e53d1..bf92dcfc 100644 --- a/examples/intrinsics/knc.h +++ b/examples/intrinsics/knc.h @@ -1,35 +1,35 @@ /* - Copyright (c) 2012-2014, Intel Corporation - All rights reserved. + Copyright (c) 2012-2014, Intel Corporation + All rights reserved. - Redistribution and use in source and binary forms, with or without - modification, are permitted provided that the following conditions are - met: + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions are +met: - * Redistributions of source code must retain the above copyright - notice, this list of conditions and the following disclaimer. + * Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. - * Redistributions in binary form must reproduce the above copyright - notice, this list of conditions and the following disclaimer in the - documentation and/or other materials provided with the distribution. + * Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. - * Neither the name of Intel Corporation nor the names of its - contributors may be used to endorse or promote products derived from - this software without specific prior written permission. + * Neither the name of Intel Corporation nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. - THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS - IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED - TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A - PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER - OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, - EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, - PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR - PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF - LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING - NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS - SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. -*/ + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS + IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED + TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A + PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ #include // INT_MIN #include @@ -66,13 +66,13 @@ #define KNC 1 extern "C" { - int printf(const unsigned char *, ...); - int puts(unsigned char *); - unsigned int putchar(unsigned int); - int fflush(void *); - uint8_t *memcpy(uint8_t *, uint8_t *, uint64_t); - uint8_t *memset(uint8_t *, uint8_t, uint64_t); - void memset_pattern16(void *, const void *, uint64_t); + int printf(const unsigned char *, ...); + int puts(unsigned char *); + unsigned int putchar(unsigned int); + int fflush(void *); + uint8_t *memcpy(uint8_t *, uint8_t *, uint64_t); + uint8_t *memset(uint8_t *, uint8_t, uint64_t); + void memset_pattern16(void *, const void *, uint64_t); } typedef float __vec1_f; @@ -86,35 +86,35 @@ struct __vec16_i32; #if 0 /* (iw) actually, this *SHOULD* be the right implementation for a - vec16_i1: this one is a class that can have a constructor (which - ISPC sometimes emits for these vectors...) This version might - not be working with embree's ISPC bindings, probably because - embree still uses the 'wrong' implementation */ +vec16_i1: this one is a class that can have a constructor (which +ISPC sometimes emits for these vectors...) This version might +not be working with embree's ISPC bindings, probably because +embree still uses the 'wrong' implementation */ typedef struct PRE_ALIGN(2) __vec16_i1 { - FORCEINLINE operator __mmask16() const { return v; } - FORCEINLINE __vec16_i1() { } - FORCEINLINE __vec16_i1(const __mmask16 &vv) : v(vv) { } - FORCEINLINE __vec16_i1(bool v0, bool v1, bool v2, bool v3, - bool v4, bool v5, bool v6, bool v7, - bool v8, bool v9, bool v10, bool v11, - bool v12, bool v13, bool v14, bool v15) { - v = ((v0 & 1) | - ((v1 & 1) << 1) | - ((v2 & 1) << 2) | - ((v3 & 1) << 3) | - ((v4 & 1) << 4) | - ((v5 & 1) << 5) | - ((v6 & 1) << 6) | - ((v7 & 1) << 7) | - ((v8 & 1) << 8) | - ((v9 & 1) << 9) | - ((v10 & 1) << 10) | - ((v11 & 1) << 11) | - ((v12 & 1) << 12) | - ((v13 & 1) << 13) | - ((v14 & 1) << 14) | - ((v15 & 1) << 15)); + FORCEINLINE operator __mmask16() const { return v; } + FORCEINLINE __vec16_i1() { } + FORCEINLINE __vec16_i1(const __mmask16 &vv) : v(vv) { } + FORCEINLINE __vec16_i1(bool v0, bool v1, bool v2, bool v3, + bool v4, bool v5, bool v6, bool v7, + bool v8, bool v9, bool v10, bool v11, + bool v12, bool v13, bool v14, bool v15) { + v = ((v0 & 1) | + ((v1 & 1) << 1) | + ((v2 & 1) << 2) | + ((v3 & 1) << 3) | + ((v4 & 1) << 4) | + ((v5 & 1) << 5) | + ((v6 & 1) << 6) | + ((v7 & 1) << 7) | + ((v8 & 1) << 8) | + ((v9 & 1) << 9) | + ((v10 & 1) << 10) | + ((v11 & 1) << 11) | + ((v12 & 1) << 12) | + ((v13 & 1) << 13) | + ((v14 & 1) << 14) | + ((v15 & 1) << 15)); } __mmask16 v; } POST_ALIGN(2) __vec16_i1; @@ -124,115 +124,115 @@ typedef __mmask16 POST_ALIGN(2) __vec16_i1; #endif typedef struct PRE_ALIGN(64) __vec16_f { - FORCEINLINE operator __m512() const { return v; } - FORCEINLINE __vec16_f() : v(_mm512_undefined_ps()) { } - FORCEINLINE __vec16_f(const __m512 &in) : v(in) {} - FORCEINLINE __vec16_f(const __vec16_f &o) : v(o.v) {} - FORCEINLINE __vec16_f& operator =(const __vec16_f &o) { v=o.v; return *this; } - FORCEINLINE __vec16_f(float v00, float v01, float v02, float v03, - float v04, float v05, float v06, float v07, - float v08, float v09, float v10, float v11, - float v12, float v13, float v14, float v15) { - v = _mm512_set_16to16_ps(v15, v14, v13, v12, v11, v10, v09, v08, v07, v06, v05, v04, v03, v02, v01, v00); - } - __m512 v; + FORCEINLINE operator __m512() const { return v; } + FORCEINLINE __vec16_f() : v(_mm512_undefined_ps()) { } + FORCEINLINE __vec16_f(const __m512 &in) : v(in) {} + FORCEINLINE __vec16_f(const __vec16_f &o) : v(o.v) {} + FORCEINLINE __vec16_f& operator =(const __vec16_f &o) { v=o.v; return *this; } + FORCEINLINE __vec16_f(float v00, float v01, float v02, float v03, + float v04, float v05, float v06, float v07, + float v08, float v09, float v10, float v11, + float v12, float v13, float v14, float v15) { + v = _mm512_set_16to16_ps(v15, v14, v13, v12, v11, v10, v09, v08, v07, v06, v05, v04, v03, v02, v01, v00); + } + __m512 v; } POST_ALIGN(64) __vec16_f; typedef struct PRE_ALIGN(64) __vec16_d { - FORCEINLINE __vec16_d() : v1(_mm512_undefined_pd()), v2(_mm512_undefined_pd()) {} - FORCEINLINE __vec16_d(const __vec16_d &o) : v1(o.v1), v2(o.v2) {} - FORCEINLINE __vec16_d(const __m512d _v1, const __m512d _v2) : v1(_v1), v2(_v2) {} - FORCEINLINE __vec16_d& operator =(const __vec16_d &o) { v1=o.v1; v2=o.v2; return *this; } - FORCEINLINE __vec16_d(double v00, double v01, double v02, double v03, - double v04, double v05, double v06, double v07, - double v08, double v09, double v10, double v11, - double v12, double v13, double v14, double v15) { - v1 = _mm512_set_8to8_pd(v15, v14, v13, v12, v11, v10, v09, v08); - v2 = _mm512_set_8to8_pd(v07, v06, v05, v04, v03, v02, v01, v00); - } - __m512d v1; - __m512d v2; + FORCEINLINE __vec16_d() : v1(_mm512_undefined_pd()), v2(_mm512_undefined_pd()) {} + FORCEINLINE __vec16_d(const __vec16_d &o) : v1(o.v1), v2(o.v2) {} + FORCEINLINE __vec16_d(const __m512d _v1, const __m512d _v2) : v1(_v1), v2(_v2) {} + FORCEINLINE __vec16_d& operator =(const __vec16_d &o) { v1=o.v1; v2=o.v2; return *this; } + FORCEINLINE __vec16_d(double v00, double v01, double v02, double v03, + double v04, double v05, double v06, double v07, + double v08, double v09, double v10, double v11, + double v12, double v13, double v14, double v15) { + v1 = _mm512_set_8to8_pd(v15, v14, v13, v12, v11, v10, v09, v08); + v2 = _mm512_set_8to8_pd(v07, v06, v05, v04, v03, v02, v01, v00); + } + __m512d v1; + __m512d v2; } POST_ALIGN(64) __vec16_d; typedef struct PRE_ALIGN(64) __vec16_i32 { - FORCEINLINE operator __m512i() const { return v; } - FORCEINLINE __vec16_i32() : v(_mm512_undefined_epi32()) {} - FORCEINLINE __vec16_i32(const int32_t &in) : v(_mm512_set1_epi32(in)) {} - FORCEINLINE __vec16_i32(const __m512i &in) : v(in) {} - FORCEINLINE __vec16_i32(const __vec16_i32 &o) : v(o.v) {} - FORCEINLINE __vec16_i32& operator =(const __vec16_i32 &o) { v=o.v; return *this; } - FORCEINLINE __vec16_i32(int32_t v00, int32_t v01, int32_t v02, int32_t v03, - int32_t v04, int32_t v05, int32_t v06, int32_t v07, - int32_t v08, int32_t v09, int32_t v10, int32_t v11, - int32_t v12, int32_t v13, int32_t v14, int32_t v15) { - v = _mm512_set_16to16_pi(v15, v14, v13, v12, v11, v10, v09, v08, v07, v06, v05, v04, v03, v02, v01, v00); - } - __m512i v; + FORCEINLINE operator __m512i() const { return v; } + FORCEINLINE __vec16_i32() : v(_mm512_undefined_epi32()) {} + FORCEINLINE __vec16_i32(const int32_t &in) : v(_mm512_set1_epi32(in)) {} + FORCEINLINE __vec16_i32(const __m512i &in) : v(in) {} + FORCEINLINE __vec16_i32(const __vec16_i32 &o) : v(o.v) {} + FORCEINLINE __vec16_i32& operator =(const __vec16_i32 &o) { v=o.v; return *this; } + FORCEINLINE __vec16_i32(int32_t v00, int32_t v01, int32_t v02, int32_t v03, + int32_t v04, int32_t v05, int32_t v06, int32_t v07, + int32_t v08, int32_t v09, int32_t v10, int32_t v11, + int32_t v12, int32_t v13, int32_t v14, int32_t v15) { + v = _mm512_set_16to16_pi(v15, v14, v13, v12, v11, v10, v09, v08, v07, v06, v05, v04, v03, v02, v01, v00); + } + __m512i v; } POST_ALIGN(64) __vec16_i32; typedef struct PRE_ALIGN(64) __vec16_i64 { - FORCEINLINE __vec16_i64() : v_lo(_mm512_undefined_epi32()), v_hi(_mm512_undefined_epi32()) {} - FORCEINLINE __vec16_i64(const __vec16_i64 &o) : v_lo(o.v_lo), v_hi(o.v_hi) {} - FORCEINLINE __vec16_i64(__m512i l, __m512i h) : v_lo(l), v_hi(h) {} - FORCEINLINE __vec16_i64& operator =(const __vec16_i64 &o) { v_lo=o.v_lo; v_hi=o.v_hi; return *this; } - FORCEINLINE __vec16_i64(int64_t v00, int64_t v01, int64_t v02, int64_t v03, - int64_t v04, int64_t v05, int64_t v06, int64_t v07, - int64_t v08, int64_t v09, int64_t v10, int64_t v11, - int64_t v12, int64_t v13, int64_t v14, int64_t v15) { - __m512i v1 = _mm512_set_8to8_epi64(v15, v14, v13, v12, v11, v10, v09, v08); - __m512i v2 = _mm512_set_8to8_epi64(v07, v06, v05, v04, v03, v02, v01, v00); - v_hi = _mm512_mask_permutevar_epi32(v_hi, 0xFF00, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - v1); - v_hi = _mm512_mask_permutevar_epi32(v_hi, 0x00FF, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - v2); - v_lo = _mm512_mask_permutevar_epi32(v_lo, 0xFF00, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - v1); - v_lo = _mm512_mask_permutevar_epi32(v_lo, 0x00FF, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - v2); - } - __m512i v_hi; - __m512i v_lo; + FORCEINLINE __vec16_i64() : v_lo(_mm512_undefined_epi32()), v_hi(_mm512_undefined_epi32()) {} + FORCEINLINE __vec16_i64(const __vec16_i64 &o) : v_lo(o.v_lo), v_hi(o.v_hi) {} + FORCEINLINE __vec16_i64(__m512i l, __m512i h) : v_lo(l), v_hi(h) {} + FORCEINLINE __vec16_i64& operator =(const __vec16_i64 &o) { v_lo=o.v_lo; v_hi=o.v_hi; return *this; } + FORCEINLINE __vec16_i64(int64_t v00, int64_t v01, int64_t v02, int64_t v03, + int64_t v04, int64_t v05, int64_t v06, int64_t v07, + int64_t v08, int64_t v09, int64_t v10, int64_t v11, + int64_t v12, int64_t v13, int64_t v14, int64_t v15) { + __m512i v1 = _mm512_set_8to8_epi64(v15, v14, v13, v12, v11, v10, v09, v08); + __m512i v2 = _mm512_set_8to8_epi64(v07, v06, v05, v04, v03, v02, v01, v00); + v_hi = _mm512_mask_permutevar_epi32(v_hi, 0xFF00, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + v1); + v_hi = _mm512_mask_permutevar_epi32(v_hi, 0x00FF, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + v2); + v_lo = _mm512_mask_permutevar_epi32(v_lo, 0xFF00, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + v1); + v_lo = _mm512_mask_permutevar_epi32(v_lo, 0x00FF, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + v2); + } + __m512i v_hi; + __m512i v_lo; } POST_ALIGN(64) __vec16_i64; template struct vec16 { - FORCEINLINE vec16() { } - FORCEINLINE vec16(T v0, T v1, T v2, T v3, T v4, T v5, T v6, T v7, - T v8, T v9, T v10, T v11, T v12, T v13, T v14, T v15) { - v[0] = v0; v[1] = v1; v[2] = v2; v[3] = v3; - v[4] = v4; v[5] = v5; v[6] = v6; v[7] = v7; - v[8] = v8; v[9] = v9; v[10] = v10; v[11] = v11; - v[12] = v12; v[13] = v13; v[14] = v14; v[15] = v15; - } - T v[16]; + FORCEINLINE vec16() { } + FORCEINLINE vec16(T v0, T v1, T v2, T v3, T v4, T v5, T v6, T v7, + T v8, T v9, T v10, T v11, T v12, T v13, T v14, T v15) { + v[0] = v0; v[1] = v1; v[2] = v2; v[3] = v3; + v[4] = v4; v[5] = v5; v[6] = v6; v[7] = v7; + v[8] = v8; v[9] = v9; v[10] = v10; v[11] = v11; + v[12] = v12; v[13] = v13; v[14] = v14; v[15] = v15; + } + T v[16]; }; PRE_ALIGN(16) struct __vec16_i8 : public vec16 { - FORCEINLINE __vec16_i8() { } - FORCEINLINE __vec16_i8(const int v0, const int v1, const int v2, const int v3, - const int v4, const int v5, const int v6, const int v7, - const int v8, const int v9, const int v10, const int v11, - const int v12, const int v13, const int v14, const int v15) - : vec16(v0, v1, v2, v3, v4, v5, v6, v7, - v8, v9, v10, v11, v12, v13, v14, v15) { } - FORCEINLINE __vec16_i8(const __vec16_i8 &o); - FORCEINLINE __vec16_i8& operator =(const __vec16_i8 &o); + FORCEINLINE __vec16_i8() { } + FORCEINLINE __vec16_i8(const int v0, const int v1, const int v2, const int v3, + const int v4, const int v5, const int v6, const int v7, + const int v8, const int v9, const int v10, const int v11, + const int v12, const int v13, const int v14, const int v15) + : vec16(v0, v1, v2, v3, v4, v5, v6, v7, + v8, v9, v10, v11, v12, v13, v14, v15) { } + FORCEINLINE __vec16_i8(const __vec16_i8 &o); + FORCEINLINE __vec16_i8& operator =(const __vec16_i8 &o); } POST_ALIGN(16); PRE_ALIGN(32) struct __vec16_i16 : public vec16 { - FORCEINLINE __vec16_i16() { } - FORCEINLINE __vec16_i16(const __vec16_i16 &o); - FORCEINLINE __vec16_i16& operator =(const __vec16_i16 &o); - FORCEINLINE __vec16_i16(int16_t v0, int16_t v1, int16_t v2, int16_t v3, - int16_t v4, int16_t v5, int16_t v6, int16_t v7, - int16_t v8, int16_t v9, int16_t v10, int16_t v11, - int16_t v12, int16_t v13, int16_t v14, int16_t v15) - : vec16(v0, v1, v2, v3, v4, v5, v6, v7, - v8, v9, v10, v11, v12, v13, v14, v15) { } + FORCEINLINE __vec16_i16() { } + FORCEINLINE __vec16_i16(const __vec16_i16 &o); + FORCEINLINE __vec16_i16& operator =(const __vec16_i16 &o); + FORCEINLINE __vec16_i16(int16_t v0, int16_t v1, int16_t v2, int16_t v3, + int16_t v4, int16_t v5, int16_t v6, int16_t v7, + int16_t v8, int16_t v9, int16_t v10, int16_t v11, + int16_t v12, int16_t v13, int16_t v14, int16_t v15) + : vec16(v0, v1, v2, v3, v4, v5, v6, v7, + v8, v9, v10, v11, v12, v13, v14, v15) { } } POST_ALIGN(32); @@ -245,8 +245,8 @@ inline std::ostream &operator<<(std::ostream &out, const __m512i &v) out << "["; for (int i=0;i<16;i++) out << (i?",":"") << std::dec << std::setw(8) << ((int*)&v)[i] << std::dec; - // out << (i?",":"") << std::hex << std::setw(8) << ((int*)&v)[i] << std::dec; - + // out << (i?",":"") << std::hex << std::setw(8) << ((int*)&v)[i] << std::dec; + out << "]" << std::flush; return out; } @@ -256,7 +256,7 @@ inline std::ostream &operator<<(std::ostream &out, const __m512 &v) out << "["; for (int i=0;i<16;i++) out << (i?",":"") << ((float*)&v)[i]; - + out << "]" << std::flush; return out; } @@ -266,8 +266,8 @@ inline std::ostream &operator<<(std::ostream &out, const __vec16_i8 &v) out << "["; for (int i=0;i<16;i++) out << (i?",":"") << std::dec << std::setw(8) << (int)((unsigned char*)&v)[i] << std::dec; - // out << (i?",":"") << std::hex << std::setw(8) << ((int*)&v)[i] << std::dec; - + // out << (i?",":"") << std::hex << std::setw(8) << ((int*)&v)[i] << std::dec; + out << "]" << std::flush; return out; } @@ -275,11 +275,11 @@ inline std::ostream &operator<<(std::ostream &out, const __vec16_i8 &v) inline std::ostream &operator<<(std::ostream &out, const __vec16_i64 &v) { out << "["; - uint32_t *ptr = (uint32_t*)&v; + uint32_t *ptr = (uint32_t*)&v; for (int i=0;i<16;i++) { -uint64_t val = (uint64_t(ptr[i])<<32)+ptr[i+16]; + uint64_t val = (uint64_t(ptr[i])<<32)+ptr[i+16]; out << (i?",":"") << ((int*)val); -} + } out << "]" << std::flush; return out; } @@ -305,66 +305,66 @@ FORCEINLINE __vec16_i8& __vec16_i8::operator=(const __vec16_i8 &o) /////////////////////////////////////////////////////////////////////////// static FORCEINLINE __vec16_i1 __movmsk(__vec16_i1 mask) { - return _mm512_kmov(mask); + return _mm512_kmov(mask); } static FORCEINLINE bool __any(__vec16_i1 mask) { - return !_mm512_kortestz(mask, mask); + return !_mm512_kortestz(mask, mask); } static FORCEINLINE bool __all(__vec16_i1 mask) { - return _mm512_kortestc(mask, mask); + return _mm512_kortestc(mask, mask); } static FORCEINLINE bool __none(__vec16_i1 mask) { - return _mm512_kortestz(mask, mask); + return _mm512_kortestz(mask, mask); } static FORCEINLINE __vec16_i1 __equal_i1(__vec16_i1 a, __vec16_i1 b) { - return _mm512_knot( _mm512_kandn(a, b)); + return _mm512_knot( _mm512_kandn(a, b)); } static FORCEINLINE __vec16_i1 __and(__vec16_i1 a, __vec16_i1 b) { - return _mm512_kand(a, b); + return _mm512_kand(a, b); } static FORCEINLINE __vec16_i1 __not(__vec16_i1 a) { - return _mm512_knot(a); + return _mm512_knot(a); } static FORCEINLINE __vec16_i1 __and_not1(__vec16_i1 a, __vec16_i1 b) { - return _mm512_kandn(a, b); + return _mm512_kandn(a, b); } static FORCEINLINE __vec16_i1 __and_not2(__vec16_i1 a, __vec16_i1 b) { - return _mm512_kandnr(a, b); + return _mm512_kandnr(a, b); } static FORCEINLINE __vec16_i1 __xor(__vec16_i1 a, __vec16_i1 b) { - return _mm512_kxor(a, b); + return _mm512_kxor(a, b); } static FORCEINLINE __vec16_i1 __xnor(__vec16_i1 a, __vec16_i1 b) { - return _mm512_kxnor(a, b); + return _mm512_kxnor(a, b); } static FORCEINLINE __vec16_i1 __or(__vec16_i1 a, __vec16_i1 b) { - return _mm512_kor(a, b); + return _mm512_kor(a, b); } static FORCEINLINE __vec16_i1 __select(__vec16_i1 mask, __vec16_i1 a, - __vec16_i1 b) { - return ((a & mask) | (b & ~mask)); - //return __or(__and(a, mask), __andnr(b, mask)); + __vec16_i1 b) { + return ((a & mask) | (b & ~mask)); + //return __or(__and(a, mask), __andnr(b, mask)); } static FORCEINLINE __vec16_i1 __select(bool cond, __vec16_i1 a, __vec16_i1 b) { - return cond ? a : b; + return cond ? a : b; } static FORCEINLINE bool __extract_element(__vec16_i1 mask, uint32_t index) { - return (mask & (1 << index)) ? true : false; + return (mask & (1 << index)) ? true : false; } @@ -382,40 +382,40 @@ static FORCEINLINE int64_t __extract_element(const __vec16_i64 &v, uint32_t inde /* -static FORCEINLINE void __insert_element(__vec16_i1 *vec, int index, - bool val) { - if (val == false) - vec->v &= ~(1 << index); - else - vec->v |= (1 << index); -} -*/ + static FORCEINLINE void __insert_element(__vec16_i1 *vec, int index, + bool val) { + if (val == false) + vec->v &= ~(1 << index); + else + vec->v |= (1 << index); + } + */ template static FORCEINLINE __vec16_i1 __load(const __vec16_i1 *p) { - const uint16_t *ptr = (const uint16_t *)p; - __vec16_i1 r; - r = *ptr; - return r; + const uint16_t *ptr = (const uint16_t *)p; + __vec16_i1 r; + r = *ptr; + return r; } template static FORCEINLINE void __store(__vec16_i1 *p, __vec16_i1 v) { - uint16_t *ptr = (uint16_t *)p; - *ptr = v; + uint16_t *ptr = (uint16_t *)p; + *ptr = v; } template static RetVecType __smear_i1(int i); template <> FORCEINLINE __vec16_i1 __smear_i1<__vec16_i1>(int i) { - return i?0xFFFF:0x0; + return i?0xFFFF:0x0; } template static RetVecType __setzero_i1(); template <> FORCEINLINE __vec16_i1 __setzero_i1<__vec16_i1>() { - return 0; + return 0; } template static RetVecType __undef_i1(); template <> FORCEINLINE __vec16_i1 __undef_i1<__vec16_i1>() { - return __vec16_i1(); + return __vec16_i1(); } /////////////////////////////////////////////////////////////////////////// @@ -424,7 +424,7 @@ template <> FORCEINLINE __vec16_i1 __undef_i1<__vec16_i1>() { /* -TODO + TODO */ @@ -435,7 +435,7 @@ TODO /* -TODO + TODO */ @@ -444,179 +444,179 @@ TODO /////////////////////////////////////////////////////////////////////////// static FORCEINLINE __vec16_i32 __add(__vec16_i32 a, __vec16_i32 b) { - return _mm512_add_epi32(a, b); + return _mm512_add_epi32(a, b); } static FORCEINLINE __vec16_i32 __sub(__vec16_i32 a, __vec16_i32 b) { - return _mm512_sub_epi32(a, b); + return _mm512_sub_epi32(a, b); } static FORCEINLINE __vec16_i32 __mul(__vec16_i32 a, __vec16_i32 b) { - return _mm512_mullo_epi32(a, b); + return _mm512_mullo_epi32(a, b); } static FORCEINLINE __vec16_i32 __udiv(__vec16_i32 a, __vec16_i32 b) { - return _mm512_div_epu32(a, b); + return _mm512_div_epu32(a, b); } static FORCEINLINE __vec16_i32 __sdiv(__vec16_i32 a, __vec16_i32 b) { - return _mm512_div_epi32(a, b); + return _mm512_div_epi32(a, b); } static FORCEINLINE __vec16_i32 __urem(__vec16_i32 a, __vec16_i32 b) { - return _mm512_rem_epu32(a, b); + return _mm512_rem_epu32(a, b); } static FORCEINLINE __vec16_i32 __srem(__vec16_i32 a, __vec16_i32 b) { - return _mm512_rem_epi32(a, b); + return _mm512_rem_epi32(a, b); } static FORCEINLINE __vec16_i32 __or(__vec16_i32 a, __vec16_i32 b) { - return _mm512_or_epi32(a, b); + return _mm512_or_epi32(a, b); } static FORCEINLINE __vec16_i32 __and(__vec16_i32 a, __vec16_i32 b) { - return _mm512_and_epi32(a, b); + return _mm512_and_epi32(a, b); } static FORCEINLINE __vec16_i32 __xor(__vec16_i32 a, __vec16_i32 b) { - return _mm512_xor_epi32(a, b); + return _mm512_xor_epi32(a, b); } static FORCEINLINE __vec16_i32 __shl(__vec16_i32 a, __vec16_i32 b) { - return _mm512_sllv_epi32(a, b); + return _mm512_sllv_epi32(a, b); } static FORCEINLINE __vec16_i32 __lshr(__vec16_i32 a, __vec16_i32 b) { - return _mm512_srlv_epi32(a, b); + return _mm512_srlv_epi32(a, b); } static FORCEINLINE __vec16_i32 __ashr(__vec16_i32 a, __vec16_i32 b) { - return _mm512_srav_epi32(a, b); + return _mm512_srav_epi32(a, b); } static FORCEINLINE __vec16_i32 __shl(__vec16_i32 a, int32_t n) { - return _mm512_slli_epi32(a, n); + return _mm512_slli_epi32(a, n); } static FORCEINLINE __vec16_i32 __lshr(__vec16_i32 a, int32_t n) { - return _mm512_srli_epi32(a, n); + return _mm512_srli_epi32(a, n); } static FORCEINLINE __vec16_i32 __ashr(__vec16_i32 a, int32_t n) { - return _mm512_srai_epi32(a, n); + return _mm512_srai_epi32(a, n); } static FORCEINLINE __vec16_i1 __equal_i32(const __vec16_i32 &a, const __vec16_i32 &b) { - return _mm512_cmpeq_epi32_mask(a, b); + return _mm512_cmpeq_epi32_mask(a, b); } static FORCEINLINE __vec16_i1 __equal_i32_and_mask(const __vec16_i32 &a, const __vec16_i32 &b, - __vec16_i1 m) { - return _mm512_mask_cmpeq_epi32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpeq_epi32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __not_equal_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmpneq_epi32_mask(a, b); + return _mm512_cmpneq_epi32_mask(a, b); } static FORCEINLINE __vec16_i1 __not_equal_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmpneq_epi32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpneq_epi32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __unsigned_less_equal_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmple_epu32_mask(a, b); + return _mm512_cmple_epu32_mask(a, b); } static FORCEINLINE __vec16_i1 __unsigned_less_equal_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmple_epu32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmple_epu32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __signed_less_equal_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmple_epi32_mask(a, b); + return _mm512_cmple_epi32_mask(a, b); } static FORCEINLINE __vec16_i1 __signed_less_equal_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmple_epi32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmple_epi32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __unsigned_greater_equal_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmpge_epu32_mask(a, b); + return _mm512_cmpge_epu32_mask(a, b); } static FORCEINLINE __vec16_i1 __unsigned_greater_equal_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmpge_epu32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpge_epu32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __signed_greater_equal_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmpge_epi32_mask(a, b); + return _mm512_cmpge_epi32_mask(a, b); } static FORCEINLINE __vec16_i1 __signed_greater_equal_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmpge_epi32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpge_epi32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __unsigned_less_than_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmplt_epu32_mask(a, b); + return _mm512_cmplt_epu32_mask(a, b); } static FORCEINLINE __vec16_i1 __unsigned_less_than_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmplt_epu32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmplt_epu32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __signed_less_than_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmplt_epi32_mask(a, b); + return _mm512_cmplt_epi32_mask(a, b); } static FORCEINLINE __vec16_i1 __signed_less_than_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmplt_epi32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmplt_epi32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __unsigned_greater_than_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmpgt_epu32_mask(a, b); + return _mm512_cmpgt_epu32_mask(a, b); } static FORCEINLINE __vec16_i1 __unsigned_greater_than_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmpgt_epu32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpgt_epu32_mask(m, a, b); } static FORCEINLINE __vec16_i1 __signed_greater_than_i32(__vec16_i32 a, __vec16_i32 b) { - return _mm512_cmpgt_epi32_mask(a, b); + return _mm512_cmpgt_epi32_mask(a, b); } static FORCEINLINE __vec16_i1 __signed_greater_than_i32_and_mask(__vec16_i32 a, __vec16_i32 b, - __vec16_i1 m) { - return _mm512_mask_cmpgt_epi32_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpgt_epi32_mask(m, a, b); } static FORCEINLINE __vec16_i32 __select(__vec16_i1 mask, - __vec16_i32 a, __vec16_i32 b) { - return _mm512_mask_mov_epi32(b.v, mask, a.v); + __vec16_i32 a, __vec16_i32 b) { + return _mm512_mask_mov_epi32(b.v, mask, a.v); } static FORCEINLINE __vec16_i32 __select(bool cond, __vec16_i32 a, __vec16_i32 b) { - return cond ? a : b; + return cond ? a : b; } static FORCEINLINE int32_t __extract_element(__vec16_i32 v, uint32_t index) { - return ((int32_t *)&v)[index]; + return ((int32_t *)&v)[index]; } static FORCEINLINE void __insert_element(__vec16_i32 *v, uint32_t index, int32_t val) { - ((int32_t *)v)[index] = val; + ((int32_t *)v)[index] = val; } template static RetVecType __smear_i32(int32_t i); template <> FORCEINLINE __vec16_i32 __smear_i32<__vec16_i32>(int32_t i) { - return _mm512_set1_epi32(i); + return _mm512_set1_epi32(i); } static const __vec16_i32 __ispc_one = __smear_i32<__vec16_i32>(1); @@ -626,17 +626,17 @@ static const __vec16_i32 __ispc_stride1(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 template static RetVecType __setzero_i32(); template <> FORCEINLINE __vec16_i32 __setzero_i32<__vec16_i32>() { - return _mm512_setzero_epi32(); + return _mm512_setzero_epi32(); } template static RetVecType __undef_i32(); template <> FORCEINLINE __vec16_i32 __undef_i32<__vec16_i32>() { - return __vec16_i32(); + return __vec16_i32(); } static FORCEINLINE __vec16_i32 __broadcast_i32(__vec16_i32 v, int index) { - int32_t val = __extract_element(v, index & 0xf); - return _mm512_set1_epi32(val); + int32_t val = __extract_element(v, index & 0xf); + return _mm512_set1_epi32(val); } static FORCEINLINE __vec16_i32 __cast_trunc(__vec16_i32, const __vec16_i64 i64) { @@ -644,388 +644,388 @@ static FORCEINLINE __vec16_i32 __cast_trunc(__vec16_i32, const __vec16_i64 i64) } static FORCEINLINE __vec16_i32 __rotate_i32(__vec16_i32 v, int index) { - __vec16_i32 idx = __smear_i32<__vec16_i32>(index); - __vec16_i32 shuffle = _mm512_and_epi32(_mm512_add_epi32(__ispc_stride1, idx), __smear_i32<__vec16_i32>(0xf)); - return _mm512_mask_permutevar_epi32(v, 0xffff, shuffle, v); + __vec16_i32 idx = __smear_i32<__vec16_i32>(index); + __vec16_i32 shuffle = _mm512_and_epi32(_mm512_add_epi32(__ispc_stride1, idx), __smear_i32<__vec16_i32>(0xf)); + return _mm512_mask_permutevar_epi32(v, 0xffff, shuffle, v); } static FORCEINLINE __vec16_i32 __shuffle_i32(__vec16_i32 v, __vec16_i32 index) { - return _mm512_mask_permutevar_epi32(v, 0xffff, index, v); + return _mm512_mask_permutevar_epi32(v, 0xffff, index, v); } template static FORCEINLINE __vec16_i32 __load(const __vec16_i32 *p) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - return _mm512_load_epi32(p); + return _mm512_load_epi32(p); #else - __vec16_i32 v; - v = _mm512_extloadunpacklo_epi32(v, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - v = _mm512_extloadunpackhi_epi32(v, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - return v; + __vec16_i32 v; + v = _mm512_extloadunpacklo_epi32(v, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + v = _mm512_extloadunpackhi_epi32(v, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + return v; #endif } template <> FORCEINLINE __vec16_i32 __load<64>(const __vec16_i32 *p) { - return _mm512_load_epi32(p); + return _mm512_load_epi32(p); } template static FORCEINLINE void __store(__vec16_i32 *p, __vec16_i32 v) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - _mm512_store_epi32(p, v); + _mm512_store_epi32(p, v); #else - _mm512_extpackstorelo_epi32(p, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_epi32((uint8_t*)p+64, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_extpackstorelo_epi32(p, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_epi32((uint8_t*)p+64, v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); #endif } template <> FORCEINLINE void __store<64>(__vec16_i32 *p, __vec16_i32 v) { - _mm512_store_epi32(p, v); + _mm512_store_epi32(p, v); } /////////////////////////////////////////////////////////////////////////// // int64 /////////////////////////////////////////////////////////////////////////// -static FORCEINLINE + static FORCEINLINE void __masked_store_i64(void *p, const __vec16_i64 &v, __vec16_i1 mask) { - __m512i v1; - __m512i v2; - v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - v.v_hi); - v1 = _mm512_mask_permutevar_epi32(v1, 0x5555, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - v.v_lo); - v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - v.v_hi); - v2 = _mm512_mask_permutevar_epi32(v2, 0x5555, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - v.v_lo); - _mm512_mask_store_epi64(p, mask, v2); - _mm512_mask_store_epi64(((uint8_t*)p)+64, mask>>8, v1); + __m512i v1; + __m512i v2; + v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + v.v_hi); + v1 = _mm512_mask_permutevar_epi32(v1, 0x5555, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + v.v_lo); + v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + v.v_hi); + v2 = _mm512_mask_permutevar_epi32(v2, 0x5555, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + v.v_lo); + _mm512_mask_store_epi64(p, mask, v2); + _mm512_mask_store_epi64(((uint8_t*)p)+64, mask>>8, v1); } static FORCEINLINE void __insert_element(__vec16_i64 *v, uint32_t index, int64_t val) { - ((int32_t *)&v->v_hi)[index] = val>>32; - ((int32_t *)&v->v_lo)[index] = val; + ((int32_t *)&v->v_hi)[index] = val>>32; + ((int32_t *)&v->v_lo)[index] = val; } template static RetVecType __setzero_i64(); template <> FORCEINLINE __vec16_i64 __setzero_i64<__vec16_i64>() { - __vec16_i64 ret; - ret.v_lo = _mm512_setzero_epi32(); - ret.v_hi = _mm512_setzero_epi32(); - return ret; + __vec16_i64 ret; + ret.v_lo = _mm512_setzero_epi32(); + ret.v_hi = _mm512_setzero_epi32(); + return ret; } template static RetVecType __undef_i64(); template <> FORCEINLINE __vec16_i64 __undef_i64<__vec16_i64>() { - return __vec16_i64(); + return __vec16_i64(); } static FORCEINLINE __vec16_i64 __add(const __vec16_i64 &a, const __vec16_i64 &b) { - __mmask16 carry = 0; - __m512i lo = _mm512_addsetc_epi32(a.v_lo, b.v_lo, &carry); - __m512i hi = _mm512_adc_epi32(a.v_hi, carry, b.v_hi, &carry); - return __vec16_i64(lo, hi); + __mmask16 carry = 0; + __m512i lo = _mm512_addsetc_epi32(a.v_lo, b.v_lo, &carry); + __m512i hi = _mm512_adc_epi32(a.v_hi, carry, b.v_hi, &carry); + return __vec16_i64(lo, hi); } static FORCEINLINE __vec16_i64 __sub(const __vec16_i64 &a, const __vec16_i64 &b) { - __mmask16 borrow = 0; - __m512i lo = _mm512_subsetb_epi32(a.v_lo, b.v_lo, &borrow); - __m512i hi = _mm512_sbb_epi32(a.v_hi, borrow, b.v_hi, &borrow); - return __vec16_i64(lo, hi); + __mmask16 borrow = 0; + __m512i lo = _mm512_subsetb_epi32(a.v_lo, b.v_lo, &borrow); + __m512i hi = _mm512_sbb_epi32(a.v_hi, borrow, b.v_hi, &borrow); + return __vec16_i64(lo, hi); } /*! 64x32 bit mul -- address computations often use a scale that we - know is 32 bits; and 32x64 is faster than 64x64 */ + know is 32 bits; and 32x64 is faster than 64x64 */ static FORCEINLINE __vec16_i64 __mul(const __vec16_i32 &a, const __vec16_i64 &b) { - return __vec16_i64(_mm512_mullo_epi32(a.v,b.v_lo), - _mm512_add_epi32(_mm512_mullo_epi32(a.v, b.v_hi), - _mm512_mulhi_epi32(a.v, b.v_lo))); + return __vec16_i64(_mm512_mullo_epi32(a.v,b.v_lo), + _mm512_add_epi32(_mm512_mullo_epi32(a.v, b.v_hi), + _mm512_mulhi_epi32(a.v, b.v_lo))); } static FORCEINLINE __vec16_i64 __mul(const __vec16_i64 &a, const __vec16_i64 &b) { - __vec16_i32 lo = _mm512_mullo_epi32(a.v_lo,b.v_lo); - __vec16_i32 hi_m1 = _mm512_mulhi_epi32(a.v_lo, b.v_lo); - __vec16_i32 hi_m2 = _mm512_mullo_epi32(a.v_hi, b.v_lo); - __vec16_i32 hi_m3 = _mm512_mullo_epi32(a.v_lo, b.v_hi); - __mmask16 carry = 0; - __vec16_i32 hi_p23 = _mm512_addsetc_epi32(hi_m2, hi_m1, &carry); - __vec16_i32 hi = _mm512_adc_epi32(hi_m3, carry, hi_p23, &carry); - return __vec16_i64(lo, hi); + __vec16_i32 lo = _mm512_mullo_epi32(a.v_lo,b.v_lo); + __vec16_i32 hi_m1 = _mm512_mulhi_epi32(a.v_lo, b.v_lo); + __vec16_i32 hi_m2 = _mm512_mullo_epi32(a.v_hi, b.v_lo); + __vec16_i32 hi_m3 = _mm512_mullo_epi32(a.v_lo, b.v_hi); + __mmask16 carry = 0; + __vec16_i32 hi_p23 = _mm512_addsetc_epi32(hi_m2, hi_m1, &carry); + __vec16_i32 hi = _mm512_adc_epi32(hi_m3, carry, hi_p23, &carry); + return __vec16_i64(lo, hi); } static FORCEINLINE __vec16_i64 __sdiv(const __vec16_i64 &a, const __vec16_i64 &b) { - __vec16_i64 ret; - for(int i=0; i<16; i++) { - int64_t dividend = __extract_element(a, i); - int64_t divisor = __extract_element(b, i); - int64_t quotient = dividend / divisor; // SVML - __insert_element(&ret, i, quotient); - } - return ret; + __vec16_i64 ret; + for(int i=0; i<16; i++) { + int64_t dividend = __extract_element(a, i); + int64_t divisor = __extract_element(b, i); + int64_t quotient = dividend / divisor; // SVML + __insert_element(&ret, i, quotient); + } + return ret; } static FORCEINLINE __vec16_i64 __udiv(const __vec16_i64 &a, const __vec16_i64 &b) { - __vec16_i64 ret; - for(int i=0; i<16; i++) { - uint64_t dividend = __extract_element(a, i); - uint64_t divisor = __extract_element(b, i); - uint64_t quotient = dividend / divisor; // SVML - __insert_element(&ret, i, quotient); - } - return ret; + __vec16_i64 ret; + for(int i=0; i<16; i++) { + uint64_t dividend = __extract_element(a, i); + uint64_t divisor = __extract_element(b, i); + uint64_t quotient = dividend / divisor; // SVML + __insert_element(&ret, i, quotient); + } + return ret; } static FORCEINLINE __vec16_i64 __or(__vec16_i64 a, __vec16_i64 b) { - return __vec16_i64(_mm512_or_epi32(a.v_lo, b.v_lo), _mm512_or_epi32(a.v_hi, b.v_hi)); + return __vec16_i64(_mm512_or_epi32(a.v_lo, b.v_lo), _mm512_or_epi32(a.v_hi, b.v_hi)); } static FORCEINLINE __vec16_i64 __and(__vec16_i64 a, __vec16_i64 b) { - return __vec16_i64(_mm512_and_epi32(a.v_lo, b.v_lo), _mm512_and_epi32(a.v_hi, b.v_hi)); + return __vec16_i64(_mm512_and_epi32(a.v_lo, b.v_lo), _mm512_and_epi32(a.v_hi, b.v_hi)); } static FORCEINLINE __vec16_i64 __xor(__vec16_i64 a, __vec16_i64 b) { - return __vec16_i64(_mm512_xor_epi32(a.v_lo, b.v_lo), _mm512_xor_epi32(a.v_hi, b.v_hi)); + return __vec16_i64(_mm512_xor_epi32(a.v_lo, b.v_lo), _mm512_xor_epi32(a.v_hi, b.v_hi)); } static FORCEINLINE __vec16_i64 __shl(__vec16_i64 a, __vec16_i64 b) { - __vec16_i32 xfer = _mm512_srlv_epi32(a.v_lo, _mm512_sub_epi32(__ispc_thirty_two, b.v_lo)); - __vec16_i32 hi = _mm512_or_epi32(_mm512_sllv_epi32(a.v_hi, b.v_lo), xfer); - __vec16_i32 lo = _mm512_sllv_epi32(a.v_lo, b.v_lo); - return __vec16_i64(lo, hi); + __vec16_i32 xfer = _mm512_srlv_epi32(a.v_lo, _mm512_sub_epi32(__ispc_thirty_two, b.v_lo)); + __vec16_i32 hi = _mm512_or_epi32(_mm512_sllv_epi32(a.v_hi, b.v_lo), xfer); + __vec16_i32 lo = _mm512_sllv_epi32(a.v_lo, b.v_lo); + return __vec16_i64(lo, hi); } static FORCEINLINE __vec16_i64 __shl(__vec16_i64 a, unsigned long long b) { __vec16_i32 hi = _mm512_or_epi32(_mm512_slli_epi32(a.v_hi, b), - _mm512_srli_epi32(a.v_lo, 32-b)); + _mm512_srli_epi32(a.v_lo, 32-b)); __vec16_i32 lo = _mm512_slli_epi32(a.v_lo, b); return __vec16_i64(lo, hi); } static FORCEINLINE __vec16_i64 __lshr(__vec16_i64 a, __vec16_i64 b) { - __vec16_i32 shift = _mm512_sub_epi32(__ispc_thirty_two, b.v_lo); - __vec16_i32 xfer = _mm512_and_epi32(_mm512_sllv_epi32(__ispc_ffffffff, shift), _mm512_sllv_epi32(a.v_hi, shift)); - //__vec16_i32 xfer = _mm512_sllv_epi32(_mm512_and_epi32(a.v_hi, - // _mm512_sub_epi32(_mm512_sllv_epi32(__ispc_one, b.v_lo), __ispc_one)), - // _mm512_sub_epi32(__ispc_thirty_two, b.v_lo)); - __vec16_i32 hi = _mm512_srlv_epi32(a.v_hi, b.v_lo); - __vec16_i32 lo = _mm512_or_epi32(xfer, _mm512_srlv_epi32(a.v_lo, b.v_lo)); - return __vec16_i64(lo, hi); + __vec16_i32 shift = _mm512_sub_epi32(__ispc_thirty_two, b.v_lo); + __vec16_i32 xfer = _mm512_and_epi32(_mm512_sllv_epi32(__ispc_ffffffff, shift), _mm512_sllv_epi32(a.v_hi, shift)); + //__vec16_i32 xfer = _mm512_sllv_epi32(_mm512_and_epi32(a.v_hi, + // _mm512_sub_epi32(_mm512_sllv_epi32(__ispc_one, b.v_lo), __ispc_one)), + // _mm512_sub_epi32(__ispc_thirty_two, b.v_lo)); + __vec16_i32 hi = _mm512_srlv_epi32(a.v_hi, b.v_lo); + __vec16_i32 lo = _mm512_or_epi32(xfer, _mm512_srlv_epi32(a.v_lo, b.v_lo)); + return __vec16_i64(lo, hi); } static FORCEINLINE __vec16_i64 __ashr(__vec16_i64 a, __vec16_i64 b) { - __vec16_i32 xfer = _mm512_sllv_epi32(_mm512_and_epi32(a.v_hi, - _mm512_sub_epi32(_mm512_sllv_epi32(__ispc_one, b.v_lo), __ispc_one)), - _mm512_sub_epi32(__ispc_thirty_two, b.v_lo)); - __vec16_i32 hi = _mm512_srav_epi32(a.v_hi, b.v_lo); - __vec16_i32 lo = _mm512_or_epi32(xfer, _mm512_srlv_epi32(a.v_lo, b.v_lo)); - return __vec16_i64(lo, hi); + __vec16_i32 xfer = _mm512_sllv_epi32(_mm512_and_epi32(a.v_hi, + _mm512_sub_epi32(_mm512_sllv_epi32(__ispc_one, b.v_lo), __ispc_one)), + _mm512_sub_epi32(__ispc_thirty_two, b.v_lo)); + __vec16_i32 hi = _mm512_srav_epi32(a.v_hi, b.v_lo); + __vec16_i32 lo = _mm512_or_epi32(xfer, _mm512_srlv_epi32(a.v_lo, b.v_lo)); + return __vec16_i64(lo, hi); } static FORCEINLINE __vec16_i64 __ashr(__vec16_i64 a, unsigned long long b) { __vec16_i32 xfer = _mm512_slli_epi32(_mm512_and_epi32(a.v_hi, - _mm512_set1_epi32((1< static RetVecType __smear_i64(const int64_t &l); template <> FORCEINLINE __vec16_i64 __smear_i64<__vec16_i64>(const int64_t &l) { - const int *i = (const int*)&l; - return __vec16_i64(_mm512_set1_epi32(i[0]), _mm512_set1_epi32(i[1])); + const int *i = (const int*)&l; + return __vec16_i64(_mm512_set1_epi32(i[0]), _mm512_set1_epi32(i[1])); } template static FORCEINLINE __vec16_i64 __load(const __vec16_i64 *p) { - __vec16_i32 v1; - __vec16_i32 v2; - const uint8_t*ptr = (const uint8_t*)p; - v2 = _mm512_extloadunpacklo_epi32(v2, ptr, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - v2 = _mm512_extloadunpackhi_epi32(v2, ptr+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - v1 = _mm512_extloadunpacklo_epi32(v1, ptr+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - v1 = _mm512_extloadunpackhi_epi32(v1, ptr+128, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - + __vec16_i32 v1; + __vec16_i32 v2; + const uint8_t*ptr = (const uint8_t*)p; + v2 = _mm512_extloadunpacklo_epi32(v2, ptr, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + v2 = _mm512_extloadunpackhi_epi32(v2, ptr+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + v1 = _mm512_extloadunpacklo_epi32(v1, ptr+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + v1 = _mm512_extloadunpackhi_epi32(v1, ptr+128, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + __vec16_i64 ret; - ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0xFF00, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - v1); - ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0x00FF, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - v2); - ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0xFF00, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - v1); - ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0x00FF, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - v2); - return ret; + ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0xFF00, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + v1); + ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0x00FF, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + v2); + ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0xFF00, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + v1); + ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0x00FF, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + v2); + return ret; } template <> FORCEINLINE __vec16_i64 __load<64>(const __vec16_i64 *p) { - __m512i v2 = _mm512_load_epi32(p); - __m512i v1 = _mm512_load_epi32(((uint8_t*)p)+64); - __vec16_i64 ret; - ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0xFF00, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - v1); - ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0x00FF, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - v2); - ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0xFF00, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - v1); - ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0x00FF, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - v2); - return ret; + __m512i v2 = _mm512_load_epi32(p); + __m512i v1 = _mm512_load_epi32(((uint8_t*)p)+64); + __vec16_i64 ret; + ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0xFF00, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + v1); + ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0x00FF, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + v2); + ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0xFF00, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + v1); + ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0x00FF, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + v2); + return ret; } template <> FORCEINLINE __vec16_i64 __load<128>(const __vec16_i64 *p) { - return __load<64>(p); + return __load<64>(p); } template static FORCEINLINE void __store(__vec16_i64 *p, __vec16_i64 v) { - __m512i v1; - __m512i v2; - v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - v.v_hi); - v1 = _mm512_mask_permutevar_epi32(v1, 0x5555, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - v.v_lo); - v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - v.v_hi); - v2 = _mm512_mask_permutevar_epi32(v2, 0x5555, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - v.v_lo); - _mm512_extpackstorelo_epi32(p, v2, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_epi32((uint8_t*)p+64, v2, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - _mm512_extpackstorelo_epi32((uint8_t*)p+64, v1, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_epi32((uint8_t*)p+128, v1, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + __m512i v1; + __m512i v2; + v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + v.v_hi); + v1 = _mm512_mask_permutevar_epi32(v1, 0x5555, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + v.v_lo); + v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + v.v_hi); + v2 = _mm512_mask_permutevar_epi32(v2, 0x5555, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + v.v_lo); + _mm512_extpackstorelo_epi32(p, v2, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_epi32((uint8_t*)p+64, v2, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_extpackstorelo_epi32((uint8_t*)p+64, v1, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_epi32((uint8_t*)p+128, v1, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); } template <> FORCEINLINE void __store<64>(__vec16_i64 *p, __vec16_i64 v) { - __m512i v1; - __m512i v2; - v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - v.v_hi); - v1 = _mm512_mask_permutevar_epi32(v1, 0x5555, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - v.v_lo); - v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - v.v_hi); - v2 = _mm512_mask_permutevar_epi32(v2, 0x5555, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - v.v_lo); - _mm512_store_epi64(p, v2); - _mm512_store_epi64(((uint8_t*)p)+64, v1); + __m512i v1; + __m512i v2; + v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + v.v_hi); + v1 = _mm512_mask_permutevar_epi32(v1, 0x5555, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + v.v_lo); + v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + v.v_hi); + v2 = _mm512_mask_permutevar_epi32(v2, 0x5555, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + v.v_lo); + _mm512_store_epi64(p, v2); + _mm512_store_epi64(((uint8_t*)p)+64, v1); } template <> FORCEINLINE void __store<128>(__vec16_i64 *p, __vec16_i64 v) { - __store<64>(p, v); + __store<64>(p, v); } /*! gather vector of 64-bit ints from addresses pointing to uniform ints (iw) WARNING: THIS CODE ONLY WORKS FOR GATHERS FROM ARRAYS OF - ***UNIFORM*** INT64's/POINTERS. (problem is that ispc doesn't - expose whether it's from array of uniform or array of varying - poitners, so in here there's no way to tell - only thing we can do - is pick one... + ***UNIFORM*** INT64's/POINTERS. (problem is that ispc doesn't + expose whether it's from array of uniform or array of varying + poitners, so in here there's no way to tell - only thing we can do + is pick one... */ static FORCEINLINE __vec16_i64 __gather_base_offsets32_i64(uint8_t *base, uint32_t scale, __vec16_i32 offsets, - __vec16_i1 mask) { + __vec16_i1 mask) { __vec16_i64 ret; ret.v_lo = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, - base, _MM_UPCONV_EPI32_NONE, scale, - _MM_HINT_NONE); + base, _MM_UPCONV_EPI32_NONE, scale, + _MM_HINT_NONE); ret.v_hi = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, - base+4, _MM_UPCONV_EPI32_NONE, scale, - _MM_HINT_NONE); + base+4, _MM_UPCONV_EPI32_NONE, scale, + _MM_HINT_NONE); return ret; } /*! gather vector of 64-bit ints from addresses pointing to uniform ints (iw) WARNING: THIS CODE ONLY WORKS FOR GATHERS FROM ARRAYS OF - ***UNIFORM*** INT64's/POINTERS. (problem is that ispc doesn't - expose whether it's from array of uniform or array of varying - poitners, so in here there's no way to tell - only thing we can do - is pick one... + ***UNIFORM*** INT64's/POINTERS. (problem is that ispc doesn't + expose whether it's from array of uniform or array of varying + poitners, so in here there's no way to tell - only thing we can do + is pick one... */ -static FORCEINLINE __vec16_i64 + static FORCEINLINE __vec16_i64 __gather64_i64(__vec16_i64 addr, __vec16_i1 mask) { - __vec16_i64 ret; + __vec16_i64 ret; - // There is no gather instruction with 64-bit offsets in KNC. - // We have to manually iterate over the upper 32 bits ;-) - __vec16_i1 still_to_do = mask; - const __vec16_i32 signed_offsets = _mm512_add_epi32(addr.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint32_t &hi32 = ((uint*)&addr.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(still_to_do,addr.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); - - void * base = (void*)((((unsigned long)hi32) << 32) + (unsigned long)(-(long)INT_MIN)); - ret.v_lo = _mm512_mask_i32extgather_epi32(ret.v_lo, match, signed_offsets, - base, _MM_UPCONV_EPI32_NONE, 1, - _MM_HINT_NONE); - ret.v_hi = _mm512_mask_i32extgather_epi32(ret.v_hi, match, signed_offsets, - base+4, _MM_UPCONV_EPI32_NONE, 1, - _MM_HINT_NONE); - - still_to_do = _mm512_kxor(match, still_to_do); - } - - return ret; + // There is no gather instruction with 64-bit offsets in KNC. + // We have to manually iterate over the upper 32 bits ;-) + __vec16_i1 still_to_do = mask; + const __vec16_i32 signed_offsets = _mm512_add_epi32(addr.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint32_t &hi32 = ((uint*)&addr.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(still_to_do,addr.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); + + void * base = (void*)((((unsigned long)hi32) << 32) + (unsigned long)(-(long)INT_MIN)); + ret.v_lo = _mm512_mask_i32extgather_epi32(ret.v_lo, match, signed_offsets, + base, _MM_UPCONV_EPI32_NONE, 1, + _MM_HINT_NONE); + ret.v_hi = _mm512_mask_i32extgather_epi32(ret.v_hi, match, signed_offsets, + base+4, _MM_UPCONV_EPI32_NONE, 1, + _MM_HINT_NONE); + + still_to_do = _mm512_kxor(match, still_to_do); + } + + return ret; } @@ -1035,149 +1035,149 @@ __gather64_i64(__vec16_i64 addr, __vec16_i1 mask) /////////////////////////////////////////////////////////////////////////// static FORCEINLINE __vec16_f __add(__vec16_f a, __vec16_f b) { - return _mm512_add_ps(a, b); + return _mm512_add_ps(a, b); } static FORCEINLINE __vec16_f __sub(__vec16_f a, __vec16_f b) { - return _mm512_sub_ps(a, b); + return _mm512_sub_ps(a, b); } static FORCEINLINE __vec16_f __mul(__vec16_f a, __vec16_f b) { - return _mm512_mul_ps(a, b); + return _mm512_mul_ps(a, b); } static FORCEINLINE __vec16_f __div(__vec16_f a, __vec16_f b) { - return _mm512_div_ps(a, b); + return _mm512_div_ps(a, b); } static FORCEINLINE __vec16_i1 __equal_float(__vec16_f a, __vec16_f b) { - return _mm512_cmpeq_ps_mask(a, b); + return _mm512_cmpeq_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __equal_float_and_mask(__vec16_f a, __vec16_f b, - __vec16_i1 m) { - return _mm512_mask_cmpeq_ps_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpeq_ps_mask(m, a, b); } static FORCEINLINE __vec16_i1 __not_equal_float(__vec16_f a, __vec16_f b) { - return _mm512_cmpneq_ps_mask(a, b); + return _mm512_cmpneq_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __not_equal_float_and_mask(__vec16_f a, __vec16_f b, - __vec16_i1 m) { - return _mm512_mask_cmpneq_ps_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpneq_ps_mask(m, a, b); } static FORCEINLINE __vec16_i1 __less_than_float(__vec16_f a, __vec16_f b) { - return _mm512_cmplt_ps_mask(a, b); + return _mm512_cmplt_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __less_than_float_and_mask(__vec16_f a, __vec16_f b, - __vec16_i1 m) { - return _mm512_mask_cmplt_ps_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmplt_ps_mask(m, a, b); } static FORCEINLINE __vec16_i1 __less_equal_float(__vec16_f a, __vec16_f b) { - return _mm512_cmple_ps_mask(a, b); + return _mm512_cmple_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __less_equal_float_and_mask(__vec16_f a, __vec16_f b, - __vec16_i1 m) { - return _mm512_mask_cmple_ps_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmple_ps_mask(m, a, b); } static FORCEINLINE __vec16_i1 __greater_than_float(__vec16_f a, __vec16_f b) { - return _mm512_cmpnle_ps_mask(a, b); + return _mm512_cmpnle_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __greater_than_float_and_mask(__vec16_f a, __vec16_f b, - __vec16_i1 m) { - return _mm512_mask_cmpnle_ps_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpnle_ps_mask(m, a, b); } static FORCEINLINE __vec16_i1 __greater_equal_float(__vec16_f a, __vec16_f b) { - return _mm512_cmpnlt_ps_mask(a, b); + return _mm512_cmpnlt_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __greater_equal_float_and_mask(__vec16_f a, __vec16_f b, - __vec16_i1 m) { - return _mm512_mask_cmpnlt_ps_mask(m, a, b); + __vec16_i1 m) { + return _mm512_mask_cmpnlt_ps_mask(m, a, b); } static FORCEINLINE __vec16_i1 __ordered_float(__vec16_f a, __vec16_f b) { - return _mm512_cmpord_ps_mask(a, b); + return _mm512_cmpord_ps_mask(a, b); } static FORCEINLINE __vec16_i1 __unordered_float(__vec16_f a, __vec16_f b) { - return _mm512_cmpunord_ps_mask(a, b); + return _mm512_cmpunord_ps_mask(a, b); } static FORCEINLINE __vec16_f __select(__vec16_i1 mask, __vec16_f a, __vec16_f b) { - return _mm512_mask_mov_ps(b, mask, a); + return _mm512_mask_mov_ps(b, mask, a); } static FORCEINLINE __vec16_f __select(bool cond, __vec16_f a, __vec16_f b) { - return cond ? a : b; + return cond ? a : b; } static FORCEINLINE float __extract_element(__vec16_f v, uint32_t index) { - return ((float *)&v)[index]; + return ((float *)&v)[index]; } static FORCEINLINE void __insert_element(__vec16_f *v, uint32_t index, float val) { - ((float *)v)[index] = val; + ((float *)v)[index] = val; } template static RetVecType __smear_float(float f); template <> FORCEINLINE __vec16_f __smear_float<__vec16_f>(float f) { - return _mm512_set_1to16_ps(f); + return _mm512_set_1to16_ps(f); } template static RetVecType __setzero_float(); template <> FORCEINLINE __vec16_f __setzero_float<__vec16_f>() { - return _mm512_setzero_ps(); + return _mm512_setzero_ps(); } template static RetVecType __undef_float(); template <> FORCEINLINE __vec16_f __undef_float<__vec16_f>() { - return __vec16_f(); + return __vec16_f(); } static FORCEINLINE __vec16_f __broadcast_float(__vec16_f v, int index) { - int32_t val = __extract_element(v, index & 0xf); - return _mm512_set1_ps(val); + int32_t val = __extract_element(v, index & 0xf); + return _mm512_set1_ps(val); } static FORCEINLINE __vec16_f __shuffle_float(__vec16_f v, __vec16_i32 index) { - return _mm512_castsi512_ps(_mm512_mask_permutevar_epi32(_mm512_castps_si512(v), 0xffff, index, _mm512_castps_si512(v))); + return _mm512_castsi512_ps(_mm512_mask_permutevar_epi32(_mm512_castps_si512(v), 0xffff, index, _mm512_castps_si512(v))); } template static FORCEINLINE __vec16_f __load(const __vec16_f *p) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - return _mm512_load_ps(p); + return _mm512_load_ps(p); #else - __vec16_f v; - v = _mm512_extloadunpacklo_ps(v, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); - v = _mm512_extloadunpackhi_ps(v, (uint8_t*)p+64, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); - return v; + __vec16_f v; + v = _mm512_extloadunpacklo_ps(v, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); + v = _mm512_extloadunpackhi_ps(v, (uint8_t*)p+64, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); + return v; #endif } template <> FORCEINLINE __vec16_f __load<64>(const __vec16_f *p) { - return _mm512_load_ps(p); + return _mm512_load_ps(p); } template static FORCEINLINE void __store(__vec16_f *p, __vec16_f v) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - _mm512_store_ps(p, v); + _mm512_store_ps(p, v); #else - _mm512_extpackstorelo_ps(p, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_ps((uint8_t*)p+64, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); + _mm512_extpackstorelo_ps(p, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_ps((uint8_t*)p+64, v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); #endif } template <> FORCEINLINE void __store<64>(__vec16_f *p, __vec16_f v) { - _mm512_store_ps(p, v); + _mm512_store_ps(p, v); } @@ -1186,241 +1186,241 @@ template <> FORCEINLINE void __store<64>(__vec16_f *p, __vec16_f v) { /////////////////////////////////////////////////////////////////////////// static FORCEINLINE __vec16_d __add(__vec16_d a, __vec16_d b) { - __vec16_d ret; - ret.v1 = _mm512_add_pd(a.v1, b.v1); - ret.v2 = _mm512_add_pd(a.v2, b.v2); - return ret; + __vec16_d ret; + ret.v1 = _mm512_add_pd(a.v1, b.v1); + ret.v2 = _mm512_add_pd(a.v2, b.v2); + return ret; } static FORCEINLINE __vec16_d __sub(__vec16_d a, __vec16_d b) { - __vec16_d ret; - ret.v1 = _mm512_sub_pd(a.v1, b.v1); - ret.v2 = _mm512_sub_pd(a.v2, b.v2); - return ret; + __vec16_d ret; + ret.v1 = _mm512_sub_pd(a.v1, b.v1); + ret.v2 = _mm512_sub_pd(a.v2, b.v2); + return ret; } static FORCEINLINE __vec16_d __mul(__vec16_d a, __vec16_d b) { - __vec16_d ret; - ret.v1 = _mm512_mul_pd(a.v1, b.v1); - ret.v2 = _mm512_mul_pd(a.v2, b.v2); - return ret; + __vec16_d ret; + ret.v1 = _mm512_mul_pd(a.v1, b.v1); + ret.v2 = _mm512_mul_pd(a.v2, b.v2); + return ret; } static FORCEINLINE __vec16_d __div(__vec16_d a, __vec16_d b) { - __vec16_d ret; - ret.v1 = _mm512_div_pd(a.v1, b.v1); - ret.v2 = _mm512_div_pd(a.v2, b.v2); - return ret; + __vec16_d ret; + ret.v1 = _mm512_div_pd(a.v1, b.v1); + ret.v2 = _mm512_div_pd(a.v2, b.v2); + return ret; } static FORCEINLINE __vec16_i1 __equal_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmpeq_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmpeq_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmpeq_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmpeq_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __equal_double_and_mask(__vec16_d a, __vec16_d b, - __vec16_i1 m) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_mask_cmpeq_pd_mask(m, a.v1, b.v1); - __vec16_i1 tmp_m = m; - ret2 = _mm512_mask_cmpeq_pd_mask(_mm512_kswapb(tmp_m,tmp_m), a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 m) { + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_mask_cmpeq_pd_mask(m, a.v1, b.v1); + __vec16_i1 tmp_m = m; + ret2 = _mm512_mask_cmpeq_pd_mask(_mm512_kswapb(tmp_m,tmp_m), a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __not_equal_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmpneq_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmpneq_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmpneq_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmpneq_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __not_equal_double_and_mask(__vec16_d a, __vec16_d b, - __vec16_i1 m) { - __vec16_i1 ret1; - __vec16_i1 ret2; - __vec16_i1 tmp_m = m; - ret1 = _mm512_mask_cmpneq_pd_mask(m, a.v1, b.v1); - ret2 = _mm512_mask_cmpneq_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 m) { + __vec16_i1 ret1; + __vec16_i1 ret2; + __vec16_i1 tmp_m = m; + ret1 = _mm512_mask_cmpneq_pd_mask(m, a.v1, b.v1); + ret2 = _mm512_mask_cmpneq_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __less_than_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmplt_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmplt_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmplt_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmplt_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __less_than_double_and_mask(__vec16_d a, __vec16_d b, - __vec16_i1 m) { - __vec16_i1 ret1; - __vec16_i1 ret2; - __vec16_i1 tmp_m = m; - ret1 = _mm512_mask_cmplt_pd_mask(m, a.v1, b.v1); - ret2 = _mm512_mask_cmplt_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 m) { + __vec16_i1 ret1; + __vec16_i1 ret2; + __vec16_i1 tmp_m = m; + ret1 = _mm512_mask_cmplt_pd_mask(m, a.v1, b.v1); + ret2 = _mm512_mask_cmplt_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __less_equal_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmple_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmple_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmple_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmple_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __less_equal_double_and_mask(__vec16_d a, __vec16_d b, - __vec16_i1 m) { - __vec16_i1 ret1; - __vec16_i1 ret2; - __vec16_i1 tmp_m = m; - ret1 = _mm512_mask_cmple_pd_mask(m, a.v1, b.v1); - ret2 = _mm512_mask_cmple_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 m) { + __vec16_i1 ret1; + __vec16_i1 ret2; + __vec16_i1 tmp_m = m; + ret1 = _mm512_mask_cmple_pd_mask(m, a.v1, b.v1); + ret2 = _mm512_mask_cmple_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __greater_than_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmpnle_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmpnle_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmpnle_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmpnle_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __greater_than_double_and_mask(__vec16_d a, __vec16_d b, - __vec16_i1 m) { - __vec16_i1 ret1; - __vec16_i1 ret2; - __vec16_i1 tmp_m = m; - ret1 = _mm512_mask_cmpnle_pd_mask(m, a.v1, b.v1); - ret2 = _mm512_mask_cmpnle_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 m) { + __vec16_i1 ret1; + __vec16_i1 ret2; + __vec16_i1 tmp_m = m; + ret1 = _mm512_mask_cmpnle_pd_mask(m, a.v1, b.v1); + ret2 = _mm512_mask_cmpnle_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __greater_equal_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmpnlt_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmpnlt_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmpnlt_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmpnlt_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __greater_equal_double_and_mask(__vec16_d a, __vec16_d b, - __vec16_i1 m) { - __vec16_i1 ret1; - __vec16_i1 ret2; - __vec16_i1 tmp_m = m; - ret1 = _mm512_mask_cmpnlt_pd_mask(m, a.v1, b.v1); - ret2 = _mm512_mask_cmpnlt_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 m) { + __vec16_i1 ret1; + __vec16_i1 ret2; + __vec16_i1 tmp_m = m; + ret1 = _mm512_mask_cmpnlt_pd_mask(m, a.v1, b.v1); + ret2 = _mm512_mask_cmpnlt_pd_mask(_mm512_kswapb(tmp_m, tmp_m), a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __ordered_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmpord_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmpord_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmpord_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmpord_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_i1 __unordered_double(__vec16_d a, __vec16_d b) { - __vec16_i1 ret1; - __vec16_i1 ret2; - ret1 = _mm512_cmpunord_pd_mask(a.v1, b.v1); - ret2 = _mm512_cmpunord_pd_mask(a.v2, b.v2); - return _mm512_kmovlhb(ret1, ret2); + __vec16_i1 ret1; + __vec16_i1 ret2; + ret1 = _mm512_cmpunord_pd_mask(a.v1, b.v1); + ret2 = _mm512_cmpunord_pd_mask(a.v2, b.v2); + return _mm512_kmovlhb(ret1, ret2); } static FORCEINLINE __vec16_d __select(__vec16_i1 mask, __vec16_d a, __vec16_d b) { - __vec16_d ret; - __vec16_i1 tmp_m = mask; - ret.v1 = _mm512_mask_mov_pd(b.v1, mask, a.v1); - ret.v2 = _mm512_mask_mov_pd(b.v2, _mm512_kswapb(tmp_m, tmp_m), a.v2); - return ret; + __vec16_d ret; + __vec16_i1 tmp_m = mask; + ret.v1 = _mm512_mask_mov_pd(b.v1, mask, a.v1); + ret.v2 = _mm512_mask_mov_pd(b.v2, _mm512_kswapb(tmp_m, tmp_m), a.v2); + return ret; } static FORCEINLINE __vec16_d __select(bool cond, __vec16_d a, __vec16_d b) { - return cond ? a : b; + return cond ? a : b; } static FORCEINLINE double __extract_element(__vec16_d v, uint32_t index) { - return ((double *)&v)[index]; + return ((double *)&v)[index]; } static FORCEINLINE void __insert_element(__vec16_d *v, uint32_t index, double val) { - ((double *)v)[index] = val; + ((double *)v)[index] = val; } template static RetVecType __smear_double(double d); template <> FORCEINLINE __vec16_d __smear_double<__vec16_d>(double d) { - __vec16_d ret; - ret.v1 = _mm512_set1_pd(d); - ret.v2 = _mm512_set1_pd(d); - return ret; + __vec16_d ret; + ret.v1 = _mm512_set1_pd(d); + ret.v2 = _mm512_set1_pd(d); + return ret; } template static RetVecType __setzero_double(); template <> FORCEINLINE __vec16_d __setzero_double<__vec16_d>() { - __vec16_d ret; - ret.v1 = _mm512_setzero_pd(); - ret.v2 = _mm512_setzero_pd(); - return ret; + __vec16_d ret; + ret.v1 = _mm512_setzero_pd(); + ret.v2 = _mm512_setzero_pd(); + return ret; } template static RetVecType __undef_double(); template <> FORCEINLINE __vec16_d __undef_double<__vec16_d>() { - return __vec16_d(); + return __vec16_d(); } static FORCEINLINE __vec16_d __broadcast_double(__vec16_d v, int index) { - __vec16_d ret; - double val = __extract_element(v, index & 0xf); - ret.v1 = _mm512_set1_pd(val); - ret.v2 = _mm512_set1_pd(val); - return ret; + __vec16_d ret; + double val = __extract_element(v, index & 0xf); + ret.v1 = _mm512_set1_pd(val); + ret.v2 = _mm512_set1_pd(val); + return ret; } template static FORCEINLINE __vec16_d __load(const __vec16_d *p) { - __vec16_d ret; - ret.v1 = _mm512_extloadunpacklo_pd(ret.v1, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - ret.v1 = _mm512_extloadunpackhi_pd(ret.v1, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - ret.v2 = _mm512_extloadunpacklo_pd(ret.v2, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - ret.v2 = _mm512_extloadunpackhi_pd(ret.v2, (uint8_t*)p+128, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - return ret; + __vec16_d ret; + ret.v1 = _mm512_extloadunpacklo_pd(ret.v1, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + ret.v1 = _mm512_extloadunpackhi_pd(ret.v1, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + ret.v2 = _mm512_extloadunpacklo_pd(ret.v2, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + ret.v2 = _mm512_extloadunpackhi_pd(ret.v2, (uint8_t*)p+128, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + return ret; } template <> FORCEINLINE __vec16_d __load<64>(const __vec16_d *p) { - __vec16_d ret; - ret.v1 = _mm512_load_pd(p); - ret.v2 = _mm512_load_pd(((uint8_t*)p)+64); - return ret; + __vec16_d ret; + ret.v1 = _mm512_load_pd(p); + ret.v2 = _mm512_load_pd(((uint8_t*)p)+64); + return ret; } template <> FORCEINLINE __vec16_d __load<128>(const __vec16_d *p) { - return __load<64>(p); + return __load<64>(p); } - + template static FORCEINLINE void __store(__vec16_d *p, __vec16_d v) { - _mm512_extpackstorelo_pd(p, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_pd((uint8_t*)p+64, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); - _mm512_extpackstorelo_pd((uint8_t*)p+64, v.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_pd((uint8_t*)p+128, v.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorelo_pd(p, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_pd((uint8_t*)p+64, v.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorelo_pd((uint8_t*)p+64, v.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_pd((uint8_t*)p+128, v.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); } template <> FORCEINLINE void __store<64>(__vec16_d *p, __vec16_d v) { - _mm512_store_pd(p, v.v1); - _mm512_store_pd(((uint8_t*)p)+64, v.v2); + _mm512_store_pd(p, v.v1); + _mm512_store_pd(((uint8_t*)p)+64, v.v2); } template <> FORCEINLINE void __store<128>(__vec16_d *p, __vec16_d v) { - __store<64>(p, v); + __store<64>(p, v); } /////////////////////////////////////////////////////////////////////////// @@ -1429,12 +1429,12 @@ template <> FORCEINLINE void __store<128>(__vec16_d *p, __vec16_d v) { static FORCEINLINE __vec16_i64 __cast_sext(const __vec16_i64 &, const __vec16_i32 &val) { - return __vec16_i64(val.v,_mm512_srai_epi32(val.v,31)); + return __vec16_i64(val.v,_mm512_srai_epi32(val.v,31)); } static FORCEINLINE __vec16_i64 __cast_zext(const __vec16_i64 &, const __vec16_i32 &val) { - return __vec16_i64(val.v, _mm512_setzero_epi32()); + return __vec16_i64(val.v, _mm512_setzero_epi32()); } static FORCEINLINE __vec16_i32 __cast_sext(const __vec16_i32 &, const __vec16_i1 &val) @@ -1446,47 +1446,47 @@ static FORCEINLINE __vec16_i32 __cast_sext(const __vec16_i32 &, const __vec16_i1 static FORCEINLINE __vec16_i32 __cast_zext(const __vec16_i32 &, const __vec16_i1 &val) { - __vec16_i32 ret = _mm512_setzero_epi32(); - __vec16_i32 one = _mm512_set1_epi32(1); - return _mm512_mask_mov_epi32(ret, val, one); + __vec16_i32 ret = _mm512_setzero_epi32(); + __vec16_i32 one = _mm512_set1_epi32(1); + return _mm512_mask_mov_epi32(ret, val, one); } static FORCEINLINE __vec16_f __cast_sitofp(__vec16_f, __vec16_i8 val) { - return _mm512_extload_ps(&val, _MM_UPCONV_PS_SINT8, _MM_BROADCAST_16X16, _MM_HINT_NONE); + return _mm512_extload_ps(&val, _MM_UPCONV_PS_SINT8, _MM_BROADCAST_16X16, _MM_HINT_NONE); } static FORCEINLINE __vec16_f __cast_sitofp(__vec16_f, __vec16_i16 val) { - return _mm512_extload_ps(&val, _MM_UPCONV_PS_SINT16, _MM_BROADCAST_16X16, _MM_HINT_NONE); + return _mm512_extload_ps(&val, _MM_UPCONV_PS_SINT16, _MM_BROADCAST_16X16, _MM_HINT_NONE); } static FORCEINLINE __vec16_f __cast_sitofp(__vec16_f, __vec16_i32 val) { - return _mm512_cvtfxpnt_round_adjustepi32_ps(val, _MM_ROUND_MODE_NEAREST, _MM_EXPADJ_NONE); + return _mm512_cvtfxpnt_round_adjustepi32_ps(val, _MM_ROUND_MODE_NEAREST, _MM_EXPADJ_NONE); } static FORCEINLINE __vec16_d __cast_sitofp(__vec16_d, __vec16_i8 val) { - __vec16_i32 vi = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT8, _MM_BROADCAST_16X16, _MM_HINT_NONE); - __vec16_d ret; - ret.v1 = _mm512_cvtepi32lo_pd(vi); - __vec16_i32 other8 = _mm512_permute4f128_epi32(vi, _MM_PERM_DCDC); - ret.v2 = _mm512_cvtepi32lo_pd(other8); - return ret; + __vec16_i32 vi = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT8, _MM_BROADCAST_16X16, _MM_HINT_NONE); + __vec16_d ret; + ret.v1 = _mm512_cvtepi32lo_pd(vi); + __vec16_i32 other8 = _mm512_permute4f128_epi32(vi, _MM_PERM_DCDC); + ret.v2 = _mm512_cvtepi32lo_pd(other8); + return ret; } static FORCEINLINE __vec16_d __cast_sitofp(__vec16_d, __vec16_i16 val) { - __vec16_i32 vi = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT16, _MM_BROADCAST_16X16, _MM_HINT_NONE); - __vec16_d ret; - ret.v1 = _mm512_cvtepi32lo_pd(vi); - __vec16_i32 other8 = _mm512_permute4f128_epi32(vi, _MM_PERM_DCDC); - ret.v2 = _mm512_cvtepi32lo_pd(other8); - return ret; + __vec16_i32 vi = _mm512_extload_epi32(&val, _MM_UPCONV_EPI32_SINT16, _MM_BROADCAST_16X16, _MM_HINT_NONE); + __vec16_d ret; + ret.v1 = _mm512_cvtepi32lo_pd(vi); + __vec16_i32 other8 = _mm512_permute4f128_epi32(vi, _MM_PERM_DCDC); + ret.v2 = _mm512_cvtepi32lo_pd(other8); + return ret; } static FORCEINLINE __vec16_d __cast_sitofp(__vec16_d, __vec16_i32 val) { - __vec16_d ret; - ret.v1 = _mm512_cvtepi32lo_pd(val); - __vec16_i32 other8 = _mm512_permute4f128_epi32(val, _MM_PERM_DCDC); - ret.v2 = _mm512_cvtepi32lo_pd(other8); - return ret; + __vec16_d ret; + ret.v1 = _mm512_cvtepi32lo_pd(val); + __vec16_i32 other8 = _mm512_permute4f128_epi32(val, _MM_PERM_DCDC); + ret.v2 = _mm512_cvtepi32lo_pd(other8); + return ret; } static FORCEINLINE __vec16_f __cast_uitofp(__vec16_f, const __vec16_i8 &v) { @@ -1508,60 +1508,60 @@ static FORCEINLINE __vec16_i32 __cast_fptoui(__vec16_i32, __vec16_f val) { } static FORCEINLINE __vec16_d __cast_fpext(__vec16_d, __vec16_f val) { - __vec16_d ret; - ret.v1 = _mm512_cvtpslo_pd(val.v); - __vec16_f other8 = _mm512_permute4f128_epi32(_mm512_castps_si512(val.v), _MM_PERM_DCDC); - ret.v2 = _mm512_cvtpslo_pd(other8); - return ret; + __vec16_d ret; + ret.v1 = _mm512_cvtpslo_pd(val.v); + __vec16_f other8 = _mm512_permute4f128_epi32(_mm512_castps_si512(val.v), _MM_PERM_DCDC); + ret.v2 = _mm512_cvtpslo_pd(other8); + return ret; } static FORCEINLINE __vec16_f __cast_fptrunc(__vec16_f, __vec16_d val) { - __m512i r0i = _mm512_castps_si512(_mm512_cvtpd_pslo(val.v2)); - __m512i r1i = _mm512_castps_si512(_mm512_cvtpd_pslo(val.v1)); - return _mm512_mask_permute4f128_epi32(r1i, 0xFF00, r0i, _MM_PERM_BABA); + __m512i r0i = _mm512_castps_si512(_mm512_cvtpd_pslo(val.v2)); + __m512i r1i = _mm512_castps_si512(_mm512_cvtpd_pslo(val.v1)); + return _mm512_mask_permute4f128_epi32(r1i, 0xFF00, r0i, _MM_PERM_BABA); } static FORCEINLINE __vec16_f __cast_bits(__vec16_f, __vec16_i32 val) { - return _mm512_castsi512_ps(val); + return _mm512_castsi512_ps(val); } static FORCEINLINE __vec16_i32 __cast_bits(__vec16_i32, __vec16_f val) { - return _mm512_castps_si512(val); + return _mm512_castps_si512(val); } static FORCEINLINE __vec16_i64 __cast_bits(__vec16_i64, __vec16_d val) { - __vec16_i64 ret; - ret.v_hi = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - val.v2); - ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0x00FF, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - val.v1); - ret.v_lo = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00, - _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), - val.v2); - ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0x00FF, - _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), - val.v1); - return ret; + __vec16_i64 ret; + ret.v_hi = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + val.v2); + ret.v_hi = _mm512_mask_permutevar_epi32(ret.v_hi, 0x00FF, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + val.v1); + ret.v_lo = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xFF00, + _mm512_set_16to16_pi(14,12,10,8,6,4,2,0,15,13,11,9,7,5,3,1), + val.v2); + ret.v_lo = _mm512_mask_permutevar_epi32(ret.v_lo, 0x00FF, + _mm512_set_16to16_pi(15,13,11,9,7,5,3,1,14,12,10,8,6,4,2,0), + val.v1); + return ret; } static FORCEINLINE __vec16_d __cast_bits(__vec16_d, __vec16_i64 val) { - __vec16_d ret; - ret.v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - val.v_hi); - ret.v2 = _mm512_mask_permutevar_epi32(ret.v2, 0x5555, - _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), - val.v_lo); - ret.v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - val.v_hi); - ret.v1 = _mm512_mask_permutevar_epi32(ret.v1, 0x5555, - _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), - val.v_lo); - return ret; + __vec16_d ret; + ret.v2 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + val.v_hi); + ret.v2 = _mm512_mask_permutevar_epi32(ret.v2, 0x5555, + _mm512_set_16to16_pi(15,15,14,14,13,13,12,12,11,11,10,10,9,9,8,8), + val.v_lo); + ret.v1 = _mm512_mask_permutevar_epi32(_mm512_undefined_epi32(), 0xAAAA, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + val.v_hi); + ret.v1 = _mm512_mask_permutevar_epi32(ret.v1, 0x5555, + _mm512_set_16to16_pi(7,7,6,6,5,5,4,4,3,3,2,2,1,1,0,0), + val.v_lo); + return ret; } /////////////////////////////////////////////////////////////////////////// @@ -1569,27 +1569,27 @@ static FORCEINLINE __vec16_d __cast_bits(__vec16_d, __vec16_i64 val) { /////////////////////////////////////////////////////////////////////////// static FORCEINLINE float __round_uniform_float(float v) { - return roundf(v); + return roundf(v); } static FORCEINLINE float __floor_uniform_float(float v) { - return floorf(v); + return floorf(v); } static FORCEINLINE float __ceil_uniform_float(float v) { - return ceilf(v); + return ceilf(v); } static FORCEINLINE double __round_uniform_double(double v) { - return round(v); + return round(v); } static FORCEINLINE double __floor_uniform_double(double v) { - return floor(v); + return floor(v); } static FORCEINLINE double __ceil_uniform_double(double v) { - return ceil(v); + return ceil(v); } static FORCEINLINE __vec16_f __round_varying_float(__vec16_f v) { @@ -1648,30 +1648,30 @@ static FORCEINLINE __vec16_i32 __min_varying_uint32(__vec16_i32 v1, __vec16_i32 // sqrt/rsqrt/rcp static FORCEINLINE float __rsqrt_uniform_float(float v) { - return 1.f / sqrtf(v); + return 1.f / sqrtf(v); } static FORCEINLINE float __rcp_uniform_float(float v) { - return 1.f / v; + return 1.f / v; } static FORCEINLINE float __sqrt_uniform_float(float v) { - return sqrtf(v); + return sqrtf(v); } static FORCEINLINE double __sqrt_uniform_double(double v) { - return sqrt(v); + return sqrt(v); } static FORCEINLINE __vec16_f __sqrt_varying_float(__vec16_f v) { - return _mm512_sqrt_ps(v); + return _mm512_sqrt_ps(v); } static FORCEINLINE __vec16_f __rcp_varying_float(__vec16_f v) { #ifdef ISPC_FAST_MATH - return _mm512_rcp23_ps(v); // Approximation with 23 bits of accuracy. + return _mm512_rcp23_ps(v); // Approximation with 23 bits of accuracy. #else - return _mm512_recip_ps(v); + return _mm512_recip_ps(v); #endif } static FORCEINLINE __vec16_d __rcp_varying_double(__vec16_d x) { @@ -1688,9 +1688,9 @@ static FORCEINLINE double __rcp_uniform_double(double v) static FORCEINLINE __vec16_f __rsqrt_varying_float(__vec16_f v) { #ifdef ISPC_FAST_MATH - return _mm512_rsqrt23_ps(v); // Approximation with 0.775ULP accuracy + return _mm512_rsqrt23_ps(v); // Approximation with 0.775ULP accuracy #else - return _mm512_invsqrt_ps(v); + return _mm512_invsqrt_ps(v); #endif } @@ -1712,20 +1712,20 @@ static FORCEINLINE double __rsqrt_uniform_double(double v) /////////////////////////////////////////////////////////////////////////// static FORCEINLINE int32_t __popcnt_int32(const __vec1_i32 mask) { - return _mm_countbits_32(mask); + return _mm_countbits_32(mask); } static FORCEINLINE int32_t __popcnt_int64(const __vec1_i64 mask) { - return _mm_countbits_64(mask); + return _mm_countbits_64(mask); } static FORCEINLINE int32_t __count_trailing_zeros_i32(const __vec1_i32 mask) { - return _mm_tzcnt_32(mask); + return _mm_tzcnt_32(mask); } static FORCEINLINE int64_t __count_trailing_zeros_i64(const __vec1_i64 mask) { - return _mm_tzcnt_64(mask); + return _mm_tzcnt_64(mask); } /////////////////////////////////////////////////////////////////////////// @@ -1749,39 +1749,39 @@ static FORCEINLINE int32_t __reduce_add_i16(__vec16_i16 v) { } static FORCEINLINE uint32_t __reduce_add_i32(__vec16_i32 v) { - return _mm512_reduce_add_epi32(v); + return _mm512_reduce_add_epi32(v); } static FORCEINLINE uint32_t __reduce_min_i32(__vec16_i32 v) { - return _mm512_reduce_min_epi32(v); + return _mm512_reduce_min_epi32(v); } static FORCEINLINE uint32_t __reduce_max_i32(__vec16_i32 v) { - return _mm512_reduce_max_epi32(v); + return _mm512_reduce_max_epi32(v); } static FORCEINLINE float __reduce_add_float(__vec16_f v) { - return _mm512_reduce_add_ps(v); + return _mm512_reduce_add_ps(v); } static FORCEINLINE float __reduce_min_float(__vec16_f v) { - return _mm512_reduce_min_ps(v); + return _mm512_reduce_min_ps(v); } static FORCEINLINE float __reduce_max_float(__vec16_f v) { - return _mm512_reduce_max_ps(v); + return _mm512_reduce_max_ps(v); } static FORCEINLINE float __reduce_add_double(__vec16_d v) { - return _mm512_reduce_add_pd(v.v1) + _mm512_reduce_add_pd(v.v2); + return _mm512_reduce_add_pd(v.v1) + _mm512_reduce_add_pd(v.v2); } static FORCEINLINE float __reduce_min_double(__vec16_d v) { - return std::min(_mm512_reduce_min_pd(v.v1), _mm512_reduce_min_pd(v.v2)); + return std::min(_mm512_reduce_min_pd(v.v1), _mm512_reduce_min_pd(v.v2)); } static FORCEINLINE float __reduce_max_double(__vec16_d v) { - return std::max(_mm512_reduce_max_pd(v.v1), _mm512_reduce_max_pd(v.v2)); + return std::max(_mm512_reduce_max_pd(v.v1), _mm512_reduce_max_pd(v.v2)); } /////////////////////////////////////////////////////////////////////////// @@ -1791,48 +1791,48 @@ static FORCEINLINE float __reduce_max_double(__vec16_d v) { // Currently, when a pseudo_gather is converted into a masked load, it has to be unaligned static FORCEINLINE __vec16_i32 __masked_load_i32(void *p, __vec16_i1 mask) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - return _mm512_mask_load_epi32(__vec16_i32(), mask, p); + return _mm512_mask_load_epi32(__vec16_i32(), mask, p); #else - __vec16_i32 tmp; - tmp.v = _mm512_mask_extloadunpacklo_epi32(tmp.v, 0xFFFF, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - tmp.v = _mm512_mask_extloadunpackhi_epi32(tmp.v, 0xFFFF, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - __vec16_i32 ret; - return _mm512_mask_mov_epi32(ret.v, mask, tmp.v); + __vec16_i32 tmp; + tmp.v = _mm512_mask_extloadunpacklo_epi32(tmp.v, 0xFFFF, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + tmp.v = _mm512_mask_extloadunpackhi_epi32(tmp.v, 0xFFFF, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + __vec16_i32 ret; + return _mm512_mask_mov_epi32(ret.v, mask, tmp.v); #endif } static FORCEINLINE __vec16_f __masked_load_float(void *p, __vec16_i1 mask) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - return _mm512_mask_load_ps(_mm512_undefined_ps(), mask,p); + return _mm512_mask_load_ps(_mm512_undefined_ps(), mask,p); #else - __vec16_f tmp; - tmp.v = _mm512_mask_extloadunpacklo_ps(tmp.v, 0xFFFF, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); - tmp.v = _mm512_mask_extloadunpackhi_ps(tmp.v, 0xFFFF, (uint8_t*)p+64, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); - __vec16_f ret; - return _mm512_mask_mov_ps(ret.v, mask, tmp.v); + __vec16_f tmp; + tmp.v = _mm512_mask_extloadunpacklo_ps(tmp.v, 0xFFFF, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); + tmp.v = _mm512_mask_extloadunpackhi_ps(tmp.v, 0xFFFF, (uint8_t*)p+64, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); + __vec16_f ret; + return _mm512_mask_mov_ps(ret.v, mask, tmp.v); #endif } static FORCEINLINE __vec16_d __masked_load_double(void *p, __vec16_i1 mask) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - __vec16_d ret; - __vec16_i1 tmp_m = mask; - tmp_m = _mm512_kswapb(tmp_m, tmp_m); - ret.v1 = _mm512_mask_load_pd(ret.v1, mask, p); - ret.v2 = _mm512_mask_load_pd(ret.v2, tmp_m, (uint8_t*)p+64); - return ret; + __vec16_d ret; + __vec16_i1 tmp_m = mask; + tmp_m = _mm512_kswapb(tmp_m, tmp_m); + ret.v1 = _mm512_mask_load_pd(ret.v1, mask, p); + ret.v2 = _mm512_mask_load_pd(ret.v2, tmp_m, (uint8_t*)p+64); + return ret; #else - __vec16_d tmp; - tmp.v1 = _mm512_mask_extloadunpacklo_pd(tmp.v1, 0xFF, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v1 = _mm512_mask_extloadunpackhi_pd(tmp.v1, 0xFF, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v2 = _mm512_mask_extloadunpacklo_pd(tmp.v2, 0xFF, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v2 = _mm512_mask_extloadunpackhi_pd(tmp.v2, 0xFF, (uint8_t*)p+128, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - __vec16_d ret; - __vec16_i1 tmp_m = mask; - tmp_m = _mm512_kswapb(tmp_m, tmp_m); - ret.v1 = _mm512_mask_mov_pd(ret.v1, mask, tmp.v1); - ret.v2 = _mm512_mask_mov_pd(ret.v2, tmp_m, tmp.v2); - return ret; + __vec16_d tmp; + tmp.v1 = _mm512_mask_extloadunpacklo_pd(tmp.v1, 0xFF, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v1 = _mm512_mask_extloadunpackhi_pd(tmp.v1, 0xFF, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v2 = _mm512_mask_extloadunpacklo_pd(tmp.v2, 0xFF, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v2 = _mm512_mask_extloadunpackhi_pd(tmp.v2, 0xFF, (uint8_t*)p+128, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + __vec16_d ret; + __vec16_i1 tmp_m = mask; + tmp_m = _mm512_kswapb(tmp_m, tmp_m); + ret.v1 = _mm512_mask_mov_pd(ret.v1, mask, tmp.v1); + ret.v2 = _mm512_mask_mov_pd(ret.v2, tmp_m, tmp.v2); + return ret; #endif } @@ -1851,8 +1851,8 @@ static FORCEINLINE __vec16_i8 __masked_load_i8(void *p, __vec16_i1 mask) { PING; __vec16_i8 ret; __vec16_i32 tmp = _mm512_mask_extload_epi32(_mm512_undefined_epi32(),mask,p, - _MM_UPCONV_EPI32_SINT8, - _MM_BROADCAST32_NONE, _MM_HINT_NONE); + _MM_UPCONV_EPI32_SINT8, + _MM_BROADCAST32_NONE, _MM_HINT_NONE); _mm512_extstore_epi32(&ret, tmp, _MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); return ret; } @@ -1862,77 +1862,78 @@ template static FORCEINLINE __vec16_i8 __load(const __vec16_i8 *p) { template static FORCEINLINE void __store(__vec16_i8 *p, __vec16_i8 v) { *p = v; } + static FORCEINLINE void __scatter_base_offsets32_i8(uint8_t *b, uint32_t scale, __vec16_i32 offsets, - __vec16_i8 val, __vec16_i1 mask) + __vec16_i8 val, __vec16_i1 mask) { __vec16_i32 tmp = _mm512_extload_epi32(&val,_MM_UPCONV_EPI32_SINT8, - _MM_BROADCAST32_NONE, _MM_HINT_NONE); + _MM_BROADCAST32_NONE, _MM_HINT_NONE); printf("__scatter_base_offsets32_i8\n"); _mm512_mask_i32extscatter_epi32(b, mask, offsets, tmp, - _MM_DOWNCONV_EPI32_SINT8, scale, - _MM_HINT_NONE); + _MM_DOWNCONV_EPI32_SINT8, scale, + _MM_HINT_NONE); } static FORCEINLINE void __masked_store_i32(void *p, __vec16_i32 val, __vec16_i1 mask) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - _mm512_mask_store_epi32(p, mask, val.v); + _mm512_mask_store_epi32(p, mask, val.v); #else - __vec16_i32 tmp; - tmp.v = _mm512_extloadunpacklo_epi32(tmp.v, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - tmp.v = _mm512_extloadunpackhi_epi32(tmp.v, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - tmp.v = _mm512_mask_mov_epi32(tmp.v, mask, val.v); - _mm512_extpackstorelo_epi32(p, tmp.v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_epi32((uint8_t*)p+64, tmp.v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + __vec16_i32 tmp; + tmp.v = _mm512_extloadunpacklo_epi32(tmp.v, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + tmp.v = _mm512_extloadunpackhi_epi32(tmp.v, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + tmp.v = _mm512_mask_mov_epi32(tmp.v, mask, val.v); + _mm512_extpackstorelo_epi32(p, tmp.v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_epi32((uint8_t*)p+64, tmp.v, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); #endif } static FORCEINLINE void __masked_store_float(void *p, __vec16_f val, - __vec16_i1 mask) { + __vec16_i1 mask) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - _mm512_mask_store_ps(p, mask, val.v); + _mm512_mask_store_ps(p, mask, val.v); #else - __vec16_f tmp; - tmp.v = _mm512_extloadunpacklo_ps(tmp.v, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); - tmp.v = _mm512_extloadunpackhi_ps(tmp.v, (uint8_t*)p+64, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); - tmp.v = _mm512_mask_mov_ps(tmp.v, mask, val.v); - _mm512_extpackstorelo_ps(p, tmp.v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_ps((uint8_t*)p+64, tmp.v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); + __vec16_f tmp; + tmp.v = _mm512_extloadunpacklo_ps(tmp.v, p, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); + tmp.v = _mm512_extloadunpackhi_ps(tmp.v, (uint8_t*)p+64, _MM_UPCONV_PS_NONE, _MM_HINT_NONE); + tmp.v = _mm512_mask_mov_ps(tmp.v, mask, val.v); + _mm512_extpackstorelo_ps(p, tmp.v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_ps((uint8_t*)p+64, tmp.v, _MM_DOWNCONV_PS_NONE, _MM_HINT_NONE); #endif } static FORCEINLINE void __masked_store_double(void *p, __vec16_d val, - __vec16_i1 mask) { + __vec16_i1 mask) { #ifdef ISPC_FORCE_ALIGNED_MEMORY - __vec16_i1 tmp_m = mask; - tmp_m = _mm512_kswapb(tmp_m, tmp_m); - _mm512_mask_store_pd(p, mask, val.v1); - _mm512_mask_store_pd((uint8_t*)p+64, tmp_m, val.v2); + __vec16_i1 tmp_m = mask; + tmp_m = _mm512_kswapb(tmp_m, tmp_m); + _mm512_mask_store_pd(p, mask, val.v1); + _mm512_mask_store_pd((uint8_t*)p+64, tmp_m, val.v2); #else - __vec16_d tmp; - __vec16_i1 tmp_m = mask; - tmp_m = _mm512_kswapb(tmp_m, tmp_m); - tmp.v1 = _mm512_extloadunpacklo_pd(tmp.v1, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v1 = _mm512_extloadunpackhi_pd(tmp.v1, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v2 = _mm512_extloadunpacklo_pd(tmp.v2, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v2 = _mm512_extloadunpackhi_pd(tmp.v2, (uint8_t*)p+128, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); - tmp.v1 = _mm512_mask_mov_pd(tmp.v1, mask, val.v1); - tmp.v2 = _mm512_mask_mov_pd(tmp.v2, tmp_m, val.v2); - _mm512_extpackstorelo_pd(p, tmp.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_pd((uint8_t*)p+64, tmp.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); - _mm512_extpackstorelo_pd((uint8_t*)p+64, tmp.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); - _mm512_extpackstorehi_pd((uint8_t*)p+128, tmp.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + __vec16_d tmp; + __vec16_i1 tmp_m = mask; + tmp_m = _mm512_kswapb(tmp_m, tmp_m); + tmp.v1 = _mm512_extloadunpacklo_pd(tmp.v1, p, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v1 = _mm512_extloadunpackhi_pd(tmp.v1, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v2 = _mm512_extloadunpacklo_pd(tmp.v2, (uint8_t*)p+64, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v2 = _mm512_extloadunpackhi_pd(tmp.v2, (uint8_t*)p+128, _MM_UPCONV_PD_NONE, _MM_HINT_NONE); + tmp.v1 = _mm512_mask_mov_pd(tmp.v1, mask, val.v1); + tmp.v2 = _mm512_mask_mov_pd(tmp.v2, tmp_m, val.v2); + _mm512_extpackstorelo_pd(p, tmp.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_pd((uint8_t*)p+64, tmp.v1, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorelo_pd((uint8_t*)p+64, tmp.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); + _mm512_extpackstorehi_pd((uint8_t*)p+128, tmp.v2, _MM_DOWNCONV_PD_NONE, _MM_HINT_NONE); #endif } static FORCEINLINE void __masked_store_blend_i32(void *p, __vec16_i32 val, - __vec16_i1 mask) { - __masked_store_i32(p, val, mask); + __vec16_i1 mask) { + __masked_store_i32(p, val, mask); } static FORCEINLINE void __masked_store_blend_float(void *p, __vec16_f val, - __vec16_i1 mask) { - __masked_store_float(p, val, mask); + __vec16_i1 mask) { + __masked_store_float(p, val, mask); } /////////////////////////////////////////////////////////////////////////// @@ -1943,72 +1944,72 @@ static FORCEINLINE void __masked_store_blend_float(void *p, __vec16_f val, static FORCEINLINE __vec16_i8 __gather_base_offsets32_i8(uint8_t *base, uint32_t scale, __vec16_i32 offsets, - __vec16_i1 mask) { - // (iw): need to temporarily store as int because gathers can only return ints. - __vec16_i32 tmp = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, base, - _MM_UPCONV_EPI32_SINT8, scale, - _MM_HINT_NONE); - // now, downconverting to chars into temporary char vector - __vec16_i8 ret; - _mm512_extstore_epi32(ret.v,tmp,_MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); - return ret; + __vec16_i1 mask) { + // (iw): need to temporarily store as int because gathers can only return ints. + __vec16_i32 tmp = _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, base, + _MM_UPCONV_EPI32_SINT8, scale, + _MM_HINT_NONE); + // now, downconverting to chars into temporary char vector + __vec16_i8 ret; + _mm512_extstore_epi32(ret.v,tmp,_MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); + return ret; } static FORCEINLINE __vec16_i32 __gather_base_offsets32_i32(uint8_t *base, uint32_t scale, __vec16_i32 offsets, - __vec16_i1 mask) { - return _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, - base, _MM_UPCONV_EPI32_NONE, scale, - _MM_HINT_NONE); + __vec16_i1 mask) { + return _mm512_mask_i32extgather_epi32(_mm512_undefined_epi32(), mask, offsets, + base, _MM_UPCONV_EPI32_NONE, scale, + _MM_HINT_NONE); } static FORCEINLINE __vec16_f __gather_base_offsets32_float(uint8_t *base, uint32_t scale, __vec16_i32 offsets, - __vec16_i1 mask) { - return _mm512_mask_i32extgather_ps(_mm512_undefined_ps(), mask, offsets, - base, _MM_UPCONV_PS_NONE, scale, - _MM_HINT_NONE); + __vec16_i1 mask) { + return _mm512_mask_i32extgather_ps(_mm512_undefined_ps(), mask, offsets, + base, _MM_UPCONV_PS_NONE, scale, + _MM_HINT_NONE); } static FORCEINLINE __vec16_d __gather_base_offsets32_double(uint8_t *base, uint32_t scale, __vec16_i32 offsets, - __vec16_i1 mask) { - __vec16_d ret; - ret.v1 = _mm512_mask_i32loextgather_pd(_mm512_undefined_pd(), mask, offsets, - base, _MM_UPCONV_PD_NONE, scale, - _MM_HINT_NONE); - __m512i shuffled_offsets = _mm512_permute4f128_epi32(offsets.v, _MM_PERM_DCDC); - ret.v2 = _mm512_mask_i32loextgather_pd(_mm512_undefined_pd(), mask, shuffled_offsets, - base, _MM_UPCONV_PD_NONE, scale, - _MM_HINT_NONE); - return ret; + __vec16_i1 mask) { + __vec16_d ret; + ret.v1 = _mm512_mask_i32loextgather_pd(_mm512_undefined_pd(), mask, offsets, + base, _MM_UPCONV_PD_NONE, scale, + _MM_HINT_NONE); + __m512i shuffled_offsets = _mm512_permute4f128_epi32(offsets.v, _MM_PERM_DCDC); + ret.v2 = _mm512_mask_i32loextgather_pd(_mm512_undefined_pd(), mask, shuffled_offsets, + base, _MM_UPCONV_PD_NONE, scale, + _MM_HINT_NONE); + return ret; } static FORCEINLINE __vec16_f __gather64_float(__vec16_i64 addr, __vec16_i1 mask) { - __vec16_f ret; + __vec16_f ret; - // There is no gather instruction with 64-bit offsets in KNC. - // We have to manually iterate over the upper 32 bits ;-) - __vec16_i1 still_to_do = mask; - const __vec16_i32 signed_offsets = _mm512_add_epi32(addr.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint &hi32 = ((uint*)&addr.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(still_to_do,addr.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); - - void * base = (void*)((((unsigned long)hi32) << 32) + (unsigned long)(-(long)INT_MIN)); + // There is no gather instruction with 64-bit offsets in KNC. + // We have to manually iterate over the upper 32 bits ;-) + __vec16_i1 still_to_do = mask; + const __vec16_i32 signed_offsets = _mm512_add_epi32(addr.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint &hi32 = ((uint*)&addr.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(still_to_do,addr.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); - ret.v = _mm512_mask_i32extgather_ps(ret.v, match, signed_offsets, - base, _MM_UPCONV_PS_NONE, 1, - _MM_HINT_NONE); - still_to_do = _mm512_kxor(match, still_to_do); - } - - return ret; + void * base = (void*)((((unsigned long)hi32) << 32) + (unsigned long)(-(long)INT_MIN)); + + ret.v = _mm512_mask_i32extgather_ps(ret.v, match, signed_offsets, + base, _MM_UPCONV_PS_NONE, 1, + _MM_HINT_NONE); + still_to_do = _mm512_kxor(match, still_to_do); + } + + return ret; } @@ -2021,183 +2022,181 @@ __gather64_float(__vec16_i64 addr, __vec16_i1 mask) static FORCEINLINE __vec16_f __gather_base_offsets64_float(uint8_t *_base, uint32_t scale, __vec16_i64 offsets, - __vec16_i1 mask) { + __vec16_i1 mask) { - const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); - // There is no gather instruction with 64-bit offsets in KNC. - // We have to manually iterate over the upper 32 bits ;-) - __vec16_i1 still_to_do = mask; - __vec16_f ret; - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); - void * base = (void*)((unsigned long)_base + - ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); - - ret = _mm512_mask_i32extgather_ps(ret, match, signed_offsets, base, - _MM_UPCONV_PS_NONE, scale, - _MM_HINT_NONE); - still_to_do = _mm512_kxor(match, still_to_do); - } + const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); + // There is no gather instruction with 64-bit offsets in KNC. + // We have to manually iterate over the upper 32 bits ;-) + __vec16_i1 still_to_do = mask; + __vec16_f ret; + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); + void * base = (void*)((unsigned long)_base + + ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); - return ret; + ret = _mm512_mask_i32extgather_ps(ret, match, signed_offsets, base, + _MM_UPCONV_PS_NONE, scale, + _MM_HINT_NONE); + still_to_do = _mm512_kxor(match, still_to_do); + } + + return ret; } static FORCEINLINE __vec16_i8 __gather_base_offsets64_i8(uint8_t *_base, uint32_t scale, __vec16_i64 offsets, - __vec16_i1 mask) + __vec16_i1 mask) { - const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); - __vec16_i1 still_to_do = mask; - __vec16_i32 tmp; - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); - - void * base = (void*)((unsigned long)_base + - ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); - tmp = _mm512_mask_i32extgather_epi32(tmp, match, signed_offsets, base, - _MM_UPCONV_EPI32_SINT8, scale, - _MM_HINT_NONE); - still_to_do = _mm512_kxor(match,still_to_do); - } - __vec16_i8 ret; - _mm512_extstore_epi32(ret.v,tmp,_MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); - return ret; + const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); + __vec16_i1 still_to_do = mask; + __vec16_i32 tmp; + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); + + void * base = (void*)((unsigned long)_base + + ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); + tmp = _mm512_mask_i32extgather_epi32(tmp, match, signed_offsets, base, + _MM_UPCONV_EPI32_SINT8, scale, + _MM_HINT_NONE); + still_to_do = _mm512_kxor(match,still_to_do); + } + __vec16_i8 ret; + _mm512_extstore_epi32(ret.v,tmp,_MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); + return ret; } static FORCEINLINE void __scatter_base_offsets64_float(uint8_t *_base, uint32_t scale, __vec16_i64 offsets, - __vec16_f value, - __vec16_i1 mask) { + __vec16_f value, + __vec16_i1 mask) { - const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); - __vec16_i1 still_to_do = mask; - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); + const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); + __vec16_i1 still_to_do = mask; + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); - void * base = (void*)((unsigned long)_base + - ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); - _mm512_mask_i32extscatter_ps(base, match, signed_offsets, - value, - _MM_DOWNCONV_PS_NONE, scale, - _MM_HINT_NONE); - still_to_do = _mm512_kxor(match,still_to_do); - } + void * base = (void*)((unsigned long)_base + + ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); + _mm512_mask_i32extscatter_ps(base, match, signed_offsets, + value, + _MM_DOWNCONV_PS_NONE, scale, + _MM_HINT_NONE); + still_to_do = _mm512_kxor(match,still_to_do); + } } static FORCEINLINE void __scatter_base_offsets64_i32(uint8_t *_base, uint32_t scale, __vec16_i64 offsets, - __vec16_i32 value, - __vec16_i1 mask) { + __vec16_i32 value, + __vec16_i1 mask) { - const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); - __vec16_i1 still_to_do = mask; - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); - - void * base = (void*)((unsigned long)_base + - ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); - _mm512_mask_i32extscatter_epi32(base, match, signed_offsets, - value, - _MM_DOWNCONV_EPI32_NONE, scale, - _MM_HINT_NONE); - still_to_do = _mm512_kxor(match,still_to_do); - } + const __vec16_i32 signed_offsets = _mm512_add_epi32(offsets.v_lo, __smear_i32<__vec16_i32>((int32_t)INT_MIN)); + __vec16_i1 still_to_do = mask; + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); + + void * base = (void*)((unsigned long)_base + + ((scale*(unsigned long)hi32) << 32) + scale*(unsigned long)(-(long)INT_MIN)); + _mm512_mask_i32extscatter_epi32(base, match, signed_offsets, + value, + _MM_DOWNCONV_EPI32_NONE, scale, + _MM_HINT_NONE); + still_to_do = _mm512_kxor(match,still_to_do); + } } static FORCEINLINE void // TODO __scatter_base_offsets64_i8(uint8_t *_base, uint32_t scale, __vec16_i64 offsets, - __vec16_i8 value, - __vec16_i1 mask) { - __vec16_i1 still_to_do = mask; + __vec16_i8 value, + __vec16_i1 mask) { + __vec16_i1 still_to_do = mask; __vec16_i32 tmp = _mm512_extload_epi32(&value, _MM_UPCONV_EPI32_SINT8, - _MM_BROADCAST32_NONE, _MM_HINT_NONE); + _MM_BROADCAST32_NONE, _MM_HINT_NONE); // PING; // _mm512_mask_extstore_epi32(p, mask, tmp, _MM_DOWNCONV_EPI32_SINT8,_MM_HINT_NONE); - while (still_to_do) { - int first_active_lane = _mm_tzcnt_32((int)still_to_do); - const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; - __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, - __smear_i32<__vec16_i32>((int32_t)hi32), - _MM_CMPINT_EQ); - - void * base = (void*)((unsigned long)_base + - ((scale*(unsigned long)hi32) << 32)); - _mm512_mask_i32extscatter_epi32(base, match, offsets.v_lo, - tmp, - _MM_DOWNCONV_EPI32_SINT8, scale, - _MM_HINT_NONE); - still_to_do = _mm512_kxor(match,still_to_do); - } + while (still_to_do) { + int first_active_lane = _mm_tzcnt_32((int)still_to_do); + const uint &hi32 = ((uint*)&offsets.v_hi)[first_active_lane]; + __vec16_i1 match = _mm512_mask_cmp_epi32_mask(mask,offsets.v_hi, + __smear_i32<__vec16_i32>((int32_t)hi32), + _MM_CMPINT_EQ); + + void * base = (void*)((unsigned long)_base + + ((scale*(unsigned long)hi32) << 32)); + _mm512_mask_i32extscatter_epi32(base, match, offsets.v_lo, + tmp, + _MM_DOWNCONV_EPI32_SINT8, scale, + _MM_HINT_NONE); + still_to_do = _mm512_kxor(match,still_to_do); + } } static FORCEINLINE __vec16_i32 __gather_base_offsets64_i32(uint8_t *_base, uint32_t scale, __vec16_i64 offsets, - __vec16_i1 mask) + __vec16_i1 mask) { - __vec16_f r = __gather_base_offsets64_float(_base,scale,offsets,mask); - return (__vec16_i32&)r; + __vec16_f r = __gather_base_offsets64_float(_base,scale,offsets,mask); + return (__vec16_i32&)r; } // scatter static FORCEINLINE void __scatter_base_offsets32_i32(uint8_t *b, uint32_t scale, __vec16_i32 offsets, - __vec16_i32 val, __vec16_i1 mask) + __vec16_i32 val, __vec16_i1 mask) { - _mm512_mask_i32extscatter_epi32(b, mask, offsets, val, - _MM_DOWNCONV_EPI32_NONE, scale, - _MM_HINT_NONE); + _mm512_mask_i32extscatter_epi32(b, mask, offsets, val, + _MM_DOWNCONV_EPI32_NONE, scale, + _MM_HINT_NONE); } static FORCEINLINE void -__scatter_base_offsets32_float(void *base, uint32_t scale, __vec16_i32 offsets, - __vec16_f val, __vec16_i1 mask) +__scatter_base_offsets32_float(void *base, uint32_t scale, __vec16_i32 offsets, + __vec16_f val, __vec16_i1 mask) { - _mm512_mask_i32extscatter_ps(base, mask, offsets, val, - _MM_DOWNCONV_PS_NONE, scale, - _MM_HINT_NONE); + _mm512_mask_i32extscatter_ps(base, mask, offsets, val, + _MM_DOWNCONV_PS_NONE, scale, + _MM_HINT_NONE); } /////////////////////////////////////////////////////////////////////////// // packed load/store /////////////////////////////////////////////////////////////////////////// -static FORCEINLINE int32_t __packed_load_active(uint32_t *p, __vec16_i32 *val, - __vec16_i1 mask) { - __vec16_i32 v = __load<64>(val); - v = _mm512_mask_extloadunpacklo_epi32(v, mask, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - v = _mm512_mask_extloadunpackhi_epi32(v, mask, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); - __store<64>(val, v); - return _mm_countbits_32(uint32_t(mask)); +static FORCEINLINE int32_t __packed_load_active(uint32_t *p, __vec16_i32 *val, __vec16_i1 mask) { + __vec16_i32 v = __load<64>(val); + v = _mm512_mask_extloadunpacklo_epi32(v, mask, p, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + v = _mm512_mask_extloadunpackhi_epi32(v, mask, (uint8_t*)p+64, _MM_UPCONV_EPI32_NONE, _MM_HINT_NONE); + __store<64>(val, v); + return _mm_countbits_32(uint32_t(mask)); } -static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec16_i32 val, - __vec16_i1 mask) { - _mm512_mask_extpackstorelo_epi32(p, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - _mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); - return _mm_countbits_32(uint32_t(mask)); +static FORCEINLINE int32_t __packed_store_active(uint32_t *p, __vec16_i32 val, __vec16_i1 mask) { + _mm512_mask_extpackstorelo_epi32(p, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + _mm512_mask_extpackstorehi_epi32((uint8_t*)p+64, mask, val, _MM_DOWNCONV_EPI32_NONE, _MM_HINT_NONE); + return _mm_countbits_32(uint32_t(mask)); } static FORCEINLINE int32_t __packed_store_active2(uint32_t *p, __vec16_i32 val, __vec16_i1 mask) @@ -2210,20 +2209,20 @@ static FORCEINLINE int32_t __packed_store_active2(uint32_t *p, __vec16_i32 val, /////////////////////////////////////////////////////////////////////////// static FORCEINLINE void __prefetch_read_uniform_1(const char *p) { - _mm_prefetch(p, _MM_HINT_T0); // prefetch into L1$ + _mm_prefetch(p, _MM_HINT_T0); // prefetch into L1$ } static FORCEINLINE void __prefetch_read_uniform_2(const char *p) { - _mm_prefetch(p, _MM_HINT_T1); // prefetch into L2$ + _mm_prefetch(p, _MM_HINT_T1); // prefetch into L2$ } static FORCEINLINE void __prefetch_read_uniform_3(const char *p) { - // There is no L3$ on KNC, don't want to pollute L2$ unecessarily + // There is no L3$ on KNC, don't want to pollute L2$ unecessarily } static FORCEINLINE void __prefetch_read_uniform_nt(const char *p) { - _mm_prefetch(p, _MM_HINT_T2); // prefetch into L2$ with non-temporal hint - // _mm_prefetch(p, _MM_HINT_NTA); // prefetch into L1$ with non-temporal hint + _mm_prefetch(p, _MM_HINT_T2); // prefetch into L2$ with non-temporal hint + // _mm_prefetch(p, _MM_HINT_NTA); // prefetch into L1$ with non-temporal hint } /////////////////////////////////////////////////////////////////////////// @@ -2232,219 +2231,219 @@ static FORCEINLINE void __prefetch_read_uniform_nt(const char *p) { static FORCEINLINE uint32_t __atomic_add(uint32_t *p, uint32_t v) { #ifdef _MSC_VER - return InterlockedAdd((LONG volatile *)p, v) - v; + return InterlockedAdd((LONG volatile *)p, v) - v; #else - return __sync_fetch_and_add(p, v); + return __sync_fetch_and_add(p, v); #endif } static FORCEINLINE uint32_t __atomic_sub(uint32_t *p, uint32_t v) { #ifdef _MSC_VER - return InterlockedAdd((LONG volatile *)p, -v) + v; + return InterlockedAdd((LONG volatile *)p, -v) + v; #else - return __sync_fetch_and_sub(p, v); + return __sync_fetch_and_sub(p, v); #endif } static FORCEINLINE uint32_t __atomic_and(uint32_t *p, uint32_t v) { #ifdef _MSC_VER - return _InterlockedAnd((LONG volatile *)p, v); + return _InterlockedAnd((LONG volatile *)p, v); #else - return __sync_fetch_and_and(p, v); + return __sync_fetch_and_and(p, v); #endif } static FORCEINLINE uint32_t __atomic_or(uint32_t *p, uint32_t v) { #ifdef _MSC_VER - return _InterlockedOr((LONG volatile *)p, v); + return _InterlockedOr((LONG volatile *)p, v); #else - return __sync_fetch_and_or(p, v); + return __sync_fetch_and_or(p, v); #endif } static FORCEINLINE uint32_t __atomic_xor(uint32_t *p, uint32_t v) { #ifdef _MSC_VER - return _InterlockedXor((LONG volatile *)p, v); + return _InterlockedXor((LONG volatile *)p, v); #else - return __sync_fetch_and_xor(p, v); + return __sync_fetch_and_xor(p, v); #endif } static FORCEINLINE uint32_t __atomic_min(uint32_t *p, uint32_t v) { - int32_t old, min; - do { - old = *((volatile int32_t *)p); - min = (old < (int32_t)v) ? old : (int32_t)v; + int32_t old, min; + do { + old = *((volatile int32_t *)p); + min = (old < (int32_t)v) ? old : (int32_t)v; #ifdef _MSC_VER - } while (InterlockedCompareExchange((LONG volatile *)p, min, old) != old); + } while (InterlockedCompareExchange((LONG volatile *)p, min, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, min) == false); + } while (__sync_bool_compare_and_swap(p, old, min) == false); #endif - return old; + return old; } static FORCEINLINE uint32_t __atomic_max(uint32_t *p, uint32_t v) { - int32_t old, max; - do { - old = *((volatile int32_t *)p); - max = (old > (int32_t)v) ? old : (int32_t)v; + int32_t old, max; + do { + old = *((volatile int32_t *)p); + max = (old > (int32_t)v) ? old : (int32_t)v; #ifdef _MSC_VER - } while (InterlockedCompareExchange((LONG volatile *)p, max, old) != old); + } while (InterlockedCompareExchange((LONG volatile *)p, max, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, max) == false); + } while (__sync_bool_compare_and_swap(p, old, max) == false); #endif - return old; + return old; } static FORCEINLINE uint32_t __atomic_umin(uint32_t *p, uint32_t v) { - uint32_t old, min; - do { - old = *((volatile uint32_t *)p); - min = (old < v) ? old : v; + uint32_t old, min; + do { + old = *((volatile uint32_t *)p); + min = (old < v) ? old : v; #ifdef _MSC_VER - } while (InterlockedCompareExchange((LONG volatile *)p, min, old) != old); + } while (InterlockedCompareExchange((LONG volatile *)p, min, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, min) == false); + } while (__sync_bool_compare_and_swap(p, old, min) == false); #endif - return old; + return old; } static FORCEINLINE uint32_t __atomic_umax(uint32_t *p, uint32_t v) { - uint32_t old, max; - do { - old = *((volatile uint32_t *)p); - max = (old > v) ? old : v; + uint32_t old, max; + do { + old = *((volatile uint32_t *)p); + max = (old > v) ? old : v; #ifdef _MSC_VER - } while (InterlockedCompareExchange((LONG volatile *)p, max, old) != old); + } while (InterlockedCompareExchange((LONG volatile *)p, max, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, max) == false); + } while (__sync_bool_compare_and_swap(p, old, max) == false); #endif - return old; + return old; } static FORCEINLINE uint32_t __atomic_xchg(uint32_t *p, uint32_t v) { #ifdef _MSC_VER - return InterlockedExchange((LONG volatile *)p, v); + return InterlockedExchange((LONG volatile *)p, v); #else - return __sync_lock_test_and_set(p, v); + return __sync_lock_test_and_set(p, v); #endif } static FORCEINLINE uint32_t __atomic_cmpxchg(uint32_t *p, uint32_t cmpval, - uint32_t newval) { + uint32_t newval) { #ifdef _MSC_VER - return InterlockedCompareExchange((LONG volatile *)p, newval, cmpval); + return InterlockedCompareExchange((LONG volatile *)p, newval, cmpval); #else - return __sync_val_compare_and_swap(p, cmpval, newval); + return __sync_val_compare_and_swap(p, cmpval, newval); #endif } static FORCEINLINE uint64_t __atomic_add(uint64_t *p, uint64_t v) { #ifdef _MSC_VER - return InterlockedAdd64((LONGLONG volatile *)p, v) - v; + return InterlockedAdd64((LONGLONG volatile *)p, v) - v; #else - return __sync_fetch_and_add(p, v); + return __sync_fetch_and_add(p, v); #endif } static FORCEINLINE uint64_t __atomic_sub(uint64_t *p, uint64_t v) { #ifdef _MSC_VER - return InterlockedAdd64((LONGLONG volatile *)p, -v) + v; + return InterlockedAdd64((LONGLONG volatile *)p, -v) + v; #else - return __sync_fetch_and_sub(p, v); + return __sync_fetch_and_sub(p, v); #endif } static FORCEINLINE uint64_t __atomic_and(uint64_t *p, uint64_t v) { #ifdef _MSC_VER - return InterlockedAnd64((LONGLONG volatile *)p, v) - v; + return InterlockedAnd64((LONGLONG volatile *)p, v) - v; #else - return __sync_fetch_and_and(p, v); + return __sync_fetch_and_and(p, v); #endif } static FORCEINLINE uint64_t __atomic_or(uint64_t *p, uint64_t v) { #ifdef _MSC_VER - return InterlockedOr64((LONGLONG volatile *)p, v) - v; + return InterlockedOr64((LONGLONG volatile *)p, v) - v; #else - return __sync_fetch_and_or(p, v); + return __sync_fetch_and_or(p, v); #endif } static FORCEINLINE uint64_t __atomic_xor(uint64_t *p, uint64_t v) { #ifdef _MSC_VER - return InterlockedXor64((LONGLONG volatile *)p, v) - v; + return InterlockedXor64((LONGLONG volatile *)p, v) - v; #else - return __sync_fetch_and_xor(p, v); + return __sync_fetch_and_xor(p, v); #endif } static FORCEINLINE uint64_t __atomic_min(uint64_t *p, uint64_t v) { - int64_t old, min; - do { - old = *((volatile int64_t *)p); - min = (old < (int64_t)v) ? old : (int64_t)v; + int64_t old, min; + do { + old = *((volatile int64_t *)p); + min = (old < (int64_t)v) ? old : (int64_t)v; #ifdef _MSC_VER - } while (InterlockedCompareExchange64((LONGLONG volatile *)p, min, old) != old); + } while (InterlockedCompareExchange64((LONGLONG volatile *)p, min, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, min) == false); + } while (__sync_bool_compare_and_swap(p, old, min) == false); #endif - return old; + return old; } static FORCEINLINE uint64_t __atomic_max(uint64_t *p, uint64_t v) { - int64_t old, max; - do { - old = *((volatile int64_t *)p); - max = (old > (int64_t)v) ? old : (int64_t)v; + int64_t old, max; + do { + old = *((volatile int64_t *)p); + max = (old > (int64_t)v) ? old : (int64_t)v; #ifdef _MSC_VER - } while (InterlockedCompareExchange64((LONGLONG volatile *)p, max, old) != old); + } while (InterlockedCompareExchange64((LONGLONG volatile *)p, max, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, max) == false); + } while (__sync_bool_compare_and_swap(p, old, max) == false); #endif - return old; + return old; } static FORCEINLINE uint64_t __atomic_umin(uint64_t *p, uint64_t v) { - uint64_t old, min; - do { - old = *((volatile uint64_t *)p); - min = (old < v) ? old : v; + uint64_t old, min; + do { + old = *((volatile uint64_t *)p); + min = (old < v) ? old : v; #ifdef _MSC_VER - } while (InterlockedCompareExchange64((LONGLONG volatile *)p, min, old) != old); + } while (InterlockedCompareExchange64((LONGLONG volatile *)p, min, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, min) == false); + } while (__sync_bool_compare_and_swap(p, old, min) == false); #endif - return old; + return old; } static FORCEINLINE uint64_t __atomic_umax(uint64_t *p, uint64_t v) { - uint64_t old, max; - do { - old = *((volatile uint64_t *)p); - max = (old > v) ? old : v; + uint64_t old, max; + do { + old = *((volatile uint64_t *)p); + max = (old > v) ? old : v; #ifdef _MSC_VER - } while (InterlockedCompareExchange64((LONGLONG volatile *)p, max, old) != old); + } while (InterlockedCompareExchange64((LONGLONG volatile *)p, max, old) != old); #else - } while (__sync_bool_compare_and_swap(p, old, max) == false); + } while (__sync_bool_compare_and_swap(p, old, max) == false); #endif - return old; + return old; } static FORCEINLINE uint64_t __atomic_xchg(uint64_t *p, uint64_t v) { #ifdef _MSC_VER - return InterlockedExchange64((LONGLONG volatile *)p, v); + return InterlockedExchange64((LONGLONG volatile *)p, v); #else - return __sync_lock_test_and_set(p, v); + return __sync_lock_test_and_set(p, v); #endif } static FORCEINLINE uint64_t __atomic_cmpxchg(uint64_t *p, uint64_t cmpval, - uint64_t newval) { + uint64_t newval) { #ifdef _MSC_VER - return InterlockedCompareExchange64((LONGLONG volatile *)p, newval, cmpval); + return InterlockedCompareExchange64((LONGLONG volatile *)p, newval, cmpval); #else - return __sync_val_compare_and_swap(p, cmpval, newval); + return __sync_val_compare_and_swap(p, cmpval, newval); #endif } @@ -2456,10 +2455,10 @@ static FORCEINLINE uint64_t __clock() { uint32_t low, high; #ifdef __x86_64 __asm__ __volatile__ ("xorl %%eax,%%eax \n cpuid" - ::: "%rax", "%rbx", "%rcx", "%rdx" ); + ::: "%rax", "%rbx", "%rcx", "%rdx" ); #else __asm__ __volatile__ ("xorl %%eax,%%eax \n cpuid" - ::: "%eax", "%ebx", "%ecx", "%edx" ); + ::: "%eax", "%ebx", "%ecx", "%edx" ); #endif __asm__ __volatile__ ("rdtsc" : "=a" (low), "=d" (high)); return (uint64_t)high << 32 | low;