425 lines
12 KiB
C++

//=====================================================================
// Copyright 2006-2020 (c), Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files(the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and / or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions :
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.
//
//=====================================================================
#ifndef _SSE_VECTOR_H_
#define _SSE_VECTOR_H_
#include <xmmintrin.h>
#include <iostream>
#include <math.h>
#ifdef __linux__
class SSEVec4
#else
class __declspec(align(16)) SSEVec4
#endif
{
public:
union
{
__m128 vec128;
#ifdef __linux__
float f32[4];
#endif
};
// constructors
inline SSEVec4(){};
inline SSEVec4(float x, float y, float z, float w)
: vec128(_mm_setr_ps(x, y, z, w)){};
inline SSEVec4(__m128 vec)
: vec128(vec)
{
}
inline SSEVec4(const float* data)
: vec128(_mm_load_ps(data)){};
inline SSEVec4(float scalar)
: vec128(_mm_load1_ps(&scalar)){};
// copy and assignment
inline SSEVec4(const SSEVec4& init)
: vec128(init.vec128){};
inline const SSEVec4& operator=(const SSEVec4& lhs)
{
vec128 = lhs.vec128;
return *this;
};
// conversion to m128 type for direct use in _mm intrinsics
inline operator __m128()
{
return vec128;
};
inline operator const __m128() const
{
return vec128;
};
// indexing
#ifdef __linux__
inline const float& operator[](int i) const
{
return f32[i];
};
inline float& operator[](int i)
{
return f32[i];
};
#else
inline const float& operator[](int i) const
{
return vec128.m128_f32[i];
};
inline float& operator[](int i)
{
return vec128.m128_f32[i];
};
#endif
// addition
inline SSEVec4 operator+(const SSEVec4& rhs) const
{
return SSEVec4(_mm_add_ps(vec128, rhs.vec128));
};
inline SSEVec4& operator+=(const SSEVec4& rhs)
{
vec128 = _mm_add_ps(vec128, rhs.vec128);
return *this;
};
// multiplication
inline SSEVec4 operator*(const SSEVec4& rhs) const
{
return SSEVec4(_mm_mul_ps(vec128, rhs.vec128));
};
inline SSEVec4& operator*=(const SSEVec4& rhs)
{
vec128 = _mm_mul_ps(vec128, rhs.vec128);
return *this;
};
// scalar multiplication
//inline SSEVec4 operator*( float rhs ) const { return SSEVec4( _mm_mul_ps(vec128, _mm_load1_ps(&rhs)) ); };
//inline SSEVec4& operator*=( float rhs ) { vec128 = _mm_mul_ps(vec128, _mm_load1_ps(&rhs)); return *this; };
// subtraction
inline SSEVec4 operator-(const SSEVec4& rhs) const
{
return SSEVec4(_mm_sub_ps(vec128, rhs.vec128));
};
inline SSEVec4& operator-=(const SSEVec4& rhs)
{
vec128 = _mm_sub_ps(vec128, rhs.vec128);
return *this;
};
// division
inline SSEVec4 operator/(const SSEVec4& rhs) const
{
return SSEVec4(_mm_div_ps(vec128, rhs.vec128));
};
inline SSEVec4& operator/=(const SSEVec4& rhs)
{
vec128 = _mm_div_ps(vec128, rhs.vec128);
return *this;
};
// scalar division
inline SSEVec4 operator/(float rhs) const
{
return SSEVec4(_mm_div_ps(vec128, _mm_load1_ps(&rhs)));
};
inline SSEVec4& operator/=(float rhs)
{
vec128 = _mm_div_ps(vec128, _mm_load1_ps(&rhs));
return *this;
};
// comparison
// these return 0 or 0xffffffff in each component
inline SSEVec4 operator<(const SSEVec4& rhs) const
{
return SSEVec4(_mm_cmplt_ps(vec128, rhs.vec128));
};
inline SSEVec4 operator>(const SSEVec4& rhs) const
{
return SSEVec4(_mm_cmpgt_ps(vec128, rhs.vec128));
};
inline SSEVec4 operator<=(const SSEVec4& rhs) const
{
return SSEVec4(_mm_cmple_ps(vec128, rhs.vec128));
};
inline SSEVec4 operator>=(const SSEVec4& rhs) const
{
return SSEVec4(_mm_cmpge_ps(vec128, rhs.vec128));
};
inline SSEVec4 operator==(const SSEVec4& rhs) const
{
return SSEVec4(_mm_cmpeq_ps(vec128, rhs.vec128));
};
// bitwise operators
inline SSEVec4 operator|(const SSEVec4& rhs) const
{
return SSEVec4(_mm_or_ps(vec128, rhs.vec128));
};
inline SSEVec4 operator&(const SSEVec4& rhs) const
{
return SSEVec4(_mm_and_ps(vec128, rhs.vec128));
};
inline SSEVec4 operator^(const SSEVec4& rhs) const
{
return SSEVec4(_mm_xor_ps(vec128, rhs.vec128));
};
inline const SSEVec4& operator|=(const SSEVec4& rhs)
{
vec128 = _mm_or_ps(vec128, rhs.vec128);
return *this;
};
inline const SSEVec4& operator&=(const SSEVec4& rhs)
{
vec128 = _mm_and_ps(vec128, rhs.vec128);
return *this;
};
// for some horrible reason,there's no bitwise not instruction for SSE,
// so we have to do xor with 0xfffffff in order to fake it.
// TO get a 0xffffffff, we execute 0=0
inline SSEVec4 operator~() const
{
__m128 zero = _mm_setzero_ps();
__m128 is_true = _mm_cmpeq_ps(zero, zero);
return _mm_xor_ps(is_true, vec128);
};
};
// swizzling macros
#define VEC4_SWIZZLE(vec, x, y, z, w) SSEVec4(_mm_shuffle_ps(vec.vec128, vec.vec128, _MM_SHUFFLE(x, y, z, w)))
#define VEC4_X(vec) VEC4_SWIZZLE(vec, 0, 0, 0, 0)
#define VEC4_Y(vec) VEC4_SWIZZLE(vec, 1, 1, 1, 1)
#define VEC4_Z(vec) VEC4_SWIZZLE(vec, 2, 2, 2, 2)
#define VEC4_W(vec) VEC4_SWIZZLE(vec, 3, 3, 3, 3)
// a shuffling macro. I don't like Microsoft's, theirs is backwards
#define VEC4_SHUFFLE(x, y, z, w) _MM_SHUFFLE(w, z, y, x)
inline float SSEVec3Dot(const SSEVec4& a, const SSEVec4& b)
{
SSEVec4 tmp = a * b;
SSEVec4 sum = VEC4_Y(tmp) + VEC4_Z(tmp);
sum += tmp;
float ret;
_mm_store_ss(&ret, sum);
return ret;
};
// functions for SQRT, RSQ, RCP, MIN, MAX
#ifdef CMP_USE_RSQ_RSQR
inline SSEVec4 SSEVecQuickRCP(const SSEVec4& v) {
return SSEVec4(_mm_rcp_ps(v.vec128));
};
// RCP with newton-raphson iteration.
inline SSEVec4 SSEVecRCP(const SSEVec4& a) {
__m128 Ra0 = _mm_rcp_ps(a.vec128);
return SSEVec4(_mm_sub_ps(_mm_add_ps(Ra0, Ra0), _mm_mul_ps(_mm_mul_ps(Ra0, a.vec128), Ra0)));
};
inline SSEVec4 SSEVecQuickRSQ(const SSEVec4& v) {
return SSEVec4(_mm_rsqrt_ps(v.vec128));
};
// RSQ with newton-raphson iteration.
inline SSEVec4 SSEVecRSQ(const SSEVec4& a) {
static const SSEVec4 fvecf0pt5(0.5f);
static const SSEVec4 fvecf3pt0(3.0f);
SSEVec4 Ra0 = _mm_rsqrt_ps(a.vec128);
return (fvecf0pt5 * Ra0) * (fvecf3pt0 - (a * Ra0) * Ra0);
};
#else
inline SSEVec4 SSEVecQuickRCP(const SSEVec4& v)
{
__m128 val1 = _mm_set_ps(1.0f, 1.0f, 1.0f, 1.0f);
return SSEVec4(_mm_div_ps(val1, v.vec128)); // // return SSEVec4(_mm_rcp_ps(v.vec128));
};
// RCP with newton-raphson iteration.
inline SSEVec4 SSEVecRCP(const SSEVec4& a)
{
__m128 val1 = _mm_set_ps(1.0f, 1.0f, 1.0f, 1.0f);
__m128 Ra0 = _mm_div_ps(val1, a.vec128); // __m128 Ra0 = _mm_rcp_ps(a.vec128);
return SSEVec4(_mm_sub_ps(_mm_add_ps(Ra0, Ra0), _mm_mul_ps(_mm_mul_ps(Ra0, a.vec128), Ra0)));
};
inline SSEVec4 SSEVecQuickRSQ(const SSEVec4& v)
{
__m128 val1 = _mm_set_ps(1.0f, 1.0f, 1.0f, 1.0f);
__m128 val = _mm_sqrt_ps(v.vec128);
return SSEVec4(_mm_div_ps(val1, val)); // return SSEVec4(_mm_rsqrt_ps(v.vec128));
};
// RSQ with newton-raphson iteration.
inline SSEVec4 SSEVecRSQ(const SSEVec4& a)
{
static const SSEVec4 fvecf0pt5(0.5f);
static const SSEVec4 fvecf3pt0(3.0f);
__m128 val1 = _mm_set_ps(1.0f, 1.0f, 1.0f, 1.0f);
__m128 val = _mm_sqrt_ps(a.vec128);
SSEVec4 Ra0 = _mm_div_ps(val1, val); // SSEVec4 Ra0 = _mm_rsqrt_ps(a.vec128);
return (fvecf0pt5 * Ra0) * (fvecf3pt0 - (a * Ra0) * Ra0);
};
#endif
inline float SSEVec3Length(const SSEVec4& value)
{
float len = SSEVec3Dot(value, value);
return sqrt(len);
};
inline SSEVec4 SSEVec3Normalize(const SSEVec4& value)
{
// TODO: write a FastRSQ scalar function and use that here
float len = SSEVec3Dot(value, value);
float lenrsq = 1.0f / sqrt(len);
return value * SSEVec4(lenrsq);
};
inline SSEVec4 SSEVecSQRT(const SSEVec4& v)
{
return SSEVec4(_mm_sqrt_ps(v.vec128));
};
inline SSEVec4 SSEVecMIN(const SSEVec4& v1, const SSEVec4& v2)
{
return SSEVec4(_mm_min_ps(v1.vec128, v2.vec128));
};
inline SSEVec4 SSEVecMAX(const SSEVec4& v1, const SSEVec4& v2)
{
return SSEVec4(_mm_max_ps(v1.vec128, v2.vec128));
};
inline SSEVec4 SSEVecZero()
{
return _mm_setzero_ps();
};
// comparison functions that return an integer boolean value
// this makes direct use of the corresponding intrinsic and
// allows more efficient branching
// If you dont use these VC++.NET 2K3 will tend to dump the SSE regs
// out to the stack and suck them back in to implement a compare
inline int SSEVecIsXEqual(const SSEVec4& a, const SSEVec4& b)
{
return _mm_ucomieq_ss(a, b);
};
inline int SSEVecIsXNEqual(const SSEVec4& a, const SSEVec4& b)
{
return _mm_ucomineq_ss(a, b);
};
// these return true if the corresponding component is 0xffffffff
inline int SSEVecIsXSet(const SSEVec4& vec)
{
// HACK: -- comparisons dont work right when nans are involved
// and comparisons also return nan on true. To fix it, we're turning nan into 1.0f
// which will work when comparing against zero
return SSEVecIsXNEqual(vec & SSEVec4(1.0f), SSEVecZero());
};
inline int SSEVecIsYSet(const SSEVec4& vec)
{
return SSEVecIsXSet(VEC4_Y(vec));
};
inline int SSEVecIsZSet(const SSEVec4& vec)
{
return SSEVecIsXSet(VEC4_Z(vec));
};
inline int SSEVecIsWSet(const SSEVec4& vec)
{
return SSEVecIsXSet(VEC4_W(vec));
};
/// returns true if any component of the argument is 0xffffffff
inline int SSEVecIsAnySet(const SSEVec4& states)
{
SSEVec4 x(VEC4_X(states));
SSEVec4 y(VEC4_Y(states));
SSEVec4 z(VEC4_Z(states));
SSEVec4 w(VEC4_W(states));
SSEVec4 res = x | y | z | w;
return SSEVecIsXSet(res);
};
// loading helpers
inline SSEVec4 SSEVecLoadf(float f)
{
return _mm_load_ss(&f);
};
inline SSEVec4 SSEVecLoad2f(const float* f)
{
return _mm_loadl_pi(SSEVecZero(), (__m64*)f);
};
inline SSEVec4 SSEVecLoad3f(const float* f)
{
// load first two components and zero second two
SSEVec4 tmp = SSEVecLoad2f((float*)f);
// load the third
SSEVec4 tmp2 = SSEVecLoadf(*(f + 2));
return _mm_shuffle_ps(tmp, tmp2, VEC4_SHUFFLE(0, 1, 0, 1));
};
// storing helpers
inline void SSEVecStore2f(const SSEVec4& vec, float* pOut)
{
_mm_storel_pi((__m64*)pOut, vec.vec128);
};
inline void SSEVecStore3f(const SSEVec4& vec, float* pOut)
{
_mm_storel_pi((__m64*)pOut, vec.vec128);
_mm_store_ss((pOut + 2), VEC4_Z(vec));
};
// executes a conditional move
// target = (condition) ? val : target;
inline SSEVec4 SSEVecSelect(const SSEVec4& condition, const SSEVec4& val, const SSEVec4& target)
{
SSEVec4 res = SSEVecZero();
res = res | (val & condition);
SSEVec4 inv_cond = ~condition;
res = res | (target & inv_cond);
return res;
};
// stream output thingie for vectors
inline std::ostream& operator<<(std::ostream& out, const SSEVec4& vec)
{
out << "<" << vec[0] << "," << vec[1] << "," << vec[2] << "," << vec[3] << ">";
return out;
};
#endif