Revision control
++ // Copyright (c) 2001-2004 NovodeX AG. All rights reserved.
++ // Copyright (c) 2004-2008 AGEIA Technologies, Inc. All rights reserved.
++ // SPDX-FileCopyrightText: Copyright (c) 2008-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
++ // SPDX-License-Identifier: Apache-2.0
++
++ #ifndef PX_SIMD_HELPERS_H
++ #define PX_SIMD_HELPERS_H
++
++ #include "foundation/PxMat33.h"
++ #include "foundation/PxVecMath.h"
++ #include "foundation/PxTransform.h"
++
++ #if !PX_DOXYGEN
++ namespace physx
++ {
++ #endif
++
++ //! A padded version of PxMat33, to safely load its data using SIMD
++ class PxMat33Padded : public PxMat33
++ {
++ public:
++ explicit PX_FORCE_INLINE PxMat33Padded(const PxQuat& q)
++ {
++ using namespace aos;
++ const QuatV qV = V4LoadU(&q.x);
++ Vec3V column0V, column1V, column2V;
++ QuatGetMat33V(qV, column0V, column1V, column2V);
++ // The scalar AoS backend (COMPILE_VECTOR_INTRINSICS == 0, see PxVecMath.h) keeps Vec3V and
++ // Vec4V as distinct types, so the padded V4 stores below do not compile against it.
++ #if defined(PX_SIMD_DISABLED) || (PX_LINUX && (PX_ARM || PX_A64)) || !COMPILE_VECTOR_INTRINSICS
++ V3StoreU(column0V, column0);
++ V3StoreU(column1V, column1);
++ V3StoreU(column2V, column2);
++ #else
++ V4StoreU(column0V, &column0.x);
++ V4StoreU(column1V, &column1.x);
++ V4StoreU(column2V, &column2.x);
++ #endif
++ }
++ PX_FORCE_INLINE ~PxMat33Padded() {}
++ PX_FORCE_INLINE void operator=(const PxMat33& other)
++ {
++ column0 = other.column0;
++ column1 = other.column1;
++ column2 = other.column2;
++ }
++ PxU32 padding;
++ };
++
++ #if !PX_DOXYGEN
++ namespace aos
++ {
++ #endif
++
++ PX_FORCE_INLINE void transformKernelVec4( const FloatVArg wa, const Vec4VArg va, const Vec4VArg pa,
++ const FloatVArg wb, const Vec4VArg vb, const Vec4VArg pb,
++ FloatV& wo, Vec4V& vo, Vec4V& po)
++ {
++ wo = FSub(FMul(wa, wb), V4Dot3(va, vb));
++ vo = V4ScaleAdd(va, wb, V4ScaleAdd(vb, wa, V4Cross(va, vb)));
++
++ const Vec4V t1 = V4Scale(pb, FScaleAdd(wa, wa, FLoad(-0.5f)));
++ const Vec4V t2 = V4ScaleAdd(V4Cross(va, pb), wa, t1);
++ const Vec4V t3 = V4ScaleAdd(va, V4Dot3(va, pb), t2);
++
++ po = V4ScaleAdd(t3, FLoad(2.0f), pa);
++ }
++
++ // PT: out = a * b
++ template<const bool alignedInput, const bool alignedOutput>
++ PX_FORCE_INLINE void transformMultiply(PxTransform& out, const PxTransform& a, const PxTransform& b)
++ {
++ PX_ASSERT(!alignedInput || (size_t(&a)&15) == 0);
++ PX_ASSERT(!alignedInput || (size_t(&b)&15) == 0);
++
++ const Vec4V aPos = alignedInput ? V4LoadA(&a.p.x) : V4LoadU(&a.p.x);
++ const Vec4V aRot = alignedInput ? V4LoadA(&a.q.x) : V4LoadU(&a.q.x);
++
++ const Vec4V bPos = alignedInput ? V4LoadA(&b.p.x) : V4LoadU(&b.p.x);
++ const Vec4V bRot = alignedInput ? V4LoadA(&b.q.x) : V4LoadU(&b.q.x);
++
++ Vec4V v, p;
++ FloatV w;
++ transformKernelVec4(V4GetW(aRot), aRot, aPos, V4GetW(bRot), bRot, bPos, w, v, p);
++
++ if(alignedOutput)
++ {
++ PX_ASSERT((size_t(&out)&15) == 0);
++ V4StoreA(p, &out.p.x);
++ V4StoreA(V4SetW(v,w), &out.q.x);
++ }
++ else
++ {
++ V4StoreU(p, &out.p.x);
++ V4StoreU(V4SetW(v,w), &out.q.x);
++ }
++ }
++
++ // PT: out = a * b
++ PX_FORCE_INLINE void transformMultiply(PxTransform32& out, const PxTransform32& a, const PxTransform32& b)
++ {
++ transformMultiply<true, true>(out, a, b);
++ }
++
++ #if !PX_DOXYGEN
++ } // namespace aos
++ #endif
++
++ #if !PX_DOXYGEN
++ } // namespace physx
++ #endif
++
++ #endif