| /dports/science/gromacs/gromacs-2021.4/src/gromacs/simd/tests/ |
| H A D | bootstrap_loadstore.cpp | 85 template<typename T, typename TSimd, int simdWidth>
96 alignas(GMX_SIMD_ALIGNMENT) T src[simdWidth * 4]; in loadStoreTester()
97 alignas(GMX_SIMD_ALIGNMENT) T dst[simdWidth * 4]; in loadStoreTester()
100 T* pCopySrc = src + simdWidth + loadOffset; in loadStoreTester()
101 T* pCopyDst = dst + simdWidth + storeOffset; in loadStoreTester()
104 for (i = 0; i < simdWidth * 4; i++) in loadStoreTester()
112 for (i = 0; i < simdWidth; i++) in loadStoreTester()
118 for (i = 0; i < simdWidth * 4; i++) in loadStoreTester()
121 if (dst + i < pCopyDst || dst + i >= pCopyDst + simdWidth) in loadStoreTester()
|
| /dports/science/gromacs/gromacs-2021.4/src/gromacs/simd/ |
| H A D | simd_memory.h | 106 GMX_ASSERT((reinterpret_cast<size_t>(p) / sizeof(*p)) % simdWidth == 0, in p_()
111 p_ += simdWidth * d;
114 typename Base::difference_type operator-(SimdIterator o) { return (p_ - o.p_) / simdWidth; }
119 static constexpr int simdWidth = SimdTraits<T>::width; variable
159 GMX_ASSERT((reinterpret_cast<size_type>(begin) / sizeof(*begin)) % simdWidth == 0, in SimdArrayRef()
161 GMX_ASSERT((reinterpret_cast<size_type>(end) / sizeof(*end)) % simdWidth == 0, in SimdArrayRef()
175 size_type size() const { return (end_ - begin_) / simdWidth; } in size()
184 reference operator[](size_type n) { return reference(begin_ + n * simdWidth); }
189 reference back() const { return reference(end_ - simdWidth); } in back()
192 static constexpr int simdWidth = SimdTraits<T>::width;
[all …]
|
| /dports/math/dune-fem/dune-fem-d1c625e6d51167d0cce57de6adcef21f548a8f82/dune/fem/space/basisfunctionset/ |
| H A D | codegen.hh | 366 template < int simdWidth = 4 >
501 const size_t simdCols = simdWidth * ( numCols / simdWidth ); in evaluateCodegen()
525 for( int i=0; i< simdWidth; ++i ) in evaluateCodegen()
563 for( int i=0; i<simdWidth; ++i ) in evaluateCodegen()
573 for( int i=0; i<simdWidth; ++ i ) in evaluateCodegen()
668 const size_t simdRows = simdWidth * (numRows / simdWidth) ; in axpyCodegen()
728 for( int i=1; i<simdWidth; ++ i ) in axpyCodegen()
730 for( int i=0; i<simdWidth; ++ i ) in axpyCodegen()
891 const size_t simdNumCols = simdWidth * ( numCols / simdWidth ); in evaluateJacobiansCodegen()
1008 for( int i=0; i<simdWidth; ++i ) in evaluateJacobiansCodegen()
[all …]
|
| /dports/devel/bullet/bullet3-3.21/src/Bullet3OpenCL/RigidBody/ |
| H A D | b3GpuPgsContactSolver.h | 20 …inline int sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodie…
21 …inline int sortConstraintByBatch2(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
22 …inline int sortConstraintByBatch3(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
|
| H A D | b3GpuPgsContactSolver.cpp | 976 …int numBatches = sortConstraintByBatch3(&cpuContacts[0] + offset, n, simdWidth, csCfg.m_staticIdx,… in solveContacts()
1113 inline int b3GpuPgsContactSolver::sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int s… in sortConstraintByBatch() argument
1188 if (nCurrentBatch == simdWidth) in sortConstraintByBatch()
1234 …tSolver::sortConstraintByBatch2(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch2() argument
1238 bodyUsed2.resize(2 * simdWidth); in sortConstraintByBatch2()
1240 for (int q = 0; q < 2 * simdWidth; q++) in sortConstraintByBatch2()
1337 if (nCurrentBatch == simdWidth) in sortConstraintByBatch2()
1385 …tSolver::sortConstraintByBatch3(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
1392 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
1487 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
[all …]
|
| H A D | b3GpuPgsConstraintSolver.h | 72 …int sortConstraintByBatch3(struct b3BatchConstraint* cs, int numConstraints, int simdWidth, int st…
|
| H A D | b3GpuPgsConstraintSolver.cpp | 642 int simdWidth = numConstraints + 1; in solveGroupCacheFriendlyIterations() local
644 sortConstraintByBatch3(&batchConstraints[0], numConstraints, simdWidth, m_staticIdx, numBodies); in solveGroupCacheFriendlyIterations()
781 …::sortConstraintByBatch3(b3BatchConstraint* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
790 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
879 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
|
| /dports/devel/py-bullet3/bullet3-3.21/src/Bullet3OpenCL/RigidBody/ |
| H A D | b3GpuPgsContactSolver.h | 20 …inline int sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodie…
21 …inline int sortConstraintByBatch2(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
22 …inline int sortConstraintByBatch3(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
|
| H A D | b3GpuPgsContactSolver.cpp | 976 …int numBatches = sortConstraintByBatch3(&cpuContacts[0] + offset, n, simdWidth, csCfg.m_staticIdx,… in solveContacts()
1113 inline int b3GpuPgsContactSolver::sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int s… in sortConstraintByBatch() argument
1188 if (nCurrentBatch == simdWidth) in sortConstraintByBatch()
1234 …tSolver::sortConstraintByBatch2(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch2() argument
1238 bodyUsed2.resize(2 * simdWidth); in sortConstraintByBatch2()
1240 for (int q = 0; q < 2 * simdWidth; q++) in sortConstraintByBatch2()
1337 if (nCurrentBatch == simdWidth) in sortConstraintByBatch2()
1385 …tSolver::sortConstraintByBatch3(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
1392 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
1487 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
[all …]
|
| H A D | b3GpuPgsConstraintSolver.h | 72 …int sortConstraintByBatch3(struct b3BatchConstraint* cs, int numConstraints, int simdWidth, int st…
|
| H A D | b3GpuPgsConstraintSolver.cpp | 642 int simdWidth = numConstraints + 1; in solveGroupCacheFriendlyIterations() local
644 sortConstraintByBatch3(&batchConstraints[0], numConstraints, simdWidth, m_staticIdx, numBodies); in solveGroupCacheFriendlyIterations()
781 …::sortConstraintByBatch3(b3BatchConstraint* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
790 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
879 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
|
| /dports/devel/godot-tools/godot-3.2.3-stable/thirdparty/bullet/Bullet3OpenCL/RigidBody/ |
| H A D | b3GpuPgsContactSolver.h | 20 …inline int sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodie…
21 …inline int sortConstraintByBatch2(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
22 …inline int sortConstraintByBatch3(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
|
| H A D | b3GpuPgsContactSolver.cpp | 976 …int numBatches = sortConstraintByBatch3(&cpuContacts[0] + offset, n, simdWidth, csCfg.m_staticIdx,… in solveContacts()
1113 inline int b3GpuPgsContactSolver::sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int s… in sortConstraintByBatch() argument
1188 if (nCurrentBatch == simdWidth) in sortConstraintByBatch()
1234 …tSolver::sortConstraintByBatch2(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch2() argument
1238 bodyUsed2.resize(2 * simdWidth); in sortConstraintByBatch2()
1240 for (int q = 0; q < 2 * simdWidth; q++) in sortConstraintByBatch2()
1337 if (nCurrentBatch == simdWidth) in sortConstraintByBatch2()
1385 …tSolver::sortConstraintByBatch3(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
1392 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
1487 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
[all …]
|
| H A D | b3GpuPgsConstraintSolver.h | 72 …int sortConstraintByBatch3(struct b3BatchConstraint* cs, int numConstraints, int simdWidth, int st…
|
| H A D | b3GpuPgsConstraintSolver.cpp | 642 int simdWidth = numConstraints + 1; in solveGroupCacheFriendlyIterations() local
644 sortConstraintByBatch3(&batchConstraints[0], numConstraints, simdWidth, m_staticIdx, numBodies); in solveGroupCacheFriendlyIterations()
781 …::sortConstraintByBatch3(b3BatchConstraint* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
790 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
879 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
|
| /dports/devel/godot/godot-3.2.3-stable/thirdparty/bullet/Bullet3OpenCL/RigidBody/ |
| H A D | b3GpuPgsContactSolver.h | 20 …inline int sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodie…
21 …inline int sortConstraintByBatch2(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
22 …inline int sortConstraintByBatch3(b3Contact4* cs, int n, int simdWidth, int staticIdx, int numBodi…
|
| H A D | b3GpuPgsContactSolver.cpp | 976 …int numBatches = sortConstraintByBatch3(&cpuContacts[0] + offset, n, simdWidth, csCfg.m_staticIdx,… in solveContacts()
1113 inline int b3GpuPgsContactSolver::sortConstraintByBatch(b3Contact4* cs, int n, int simdWidth, int s… in sortConstraintByBatch() argument
1188 if (nCurrentBatch == simdWidth) in sortConstraintByBatch()
1234 …tSolver::sortConstraintByBatch2(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch2() argument
1238 bodyUsed2.resize(2 * simdWidth); in sortConstraintByBatch2()
1240 for (int q = 0; q < 2 * simdWidth; q++) in sortConstraintByBatch2()
1337 if (nCurrentBatch == simdWidth) in sortConstraintByBatch2()
1385 …tSolver::sortConstraintByBatch3(b3Contact4* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
1392 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
1487 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
[all …]
|
| H A D | b3GpuPgsConstraintSolver.h | 72 …int sortConstraintByBatch3(struct b3BatchConstraint* cs, int numConstraints, int simdWidth, int st…
|
| H A D | b3GpuPgsConstraintSolver.cpp | 642 int simdWidth = numConstraints + 1; in solveGroupCacheFriendlyIterations() local
644 sortConstraintByBatch3(&batchConstraints[0], numConstraints, simdWidth, m_staticIdx, numBodies); in solveGroupCacheFriendlyIterations()
781 …::sortConstraintByBatch3(b3BatchConstraint* cs, int numConstraints, int simdWidth, int staticIdx, … in sortConstraintByBatch3() argument
790 curUsed.resize(2 * simdWidth); in sortConstraintByBatch3()
879 if (nCurrentBatch == simdWidth) in sortConstraintByBatch3()
|
| /dports/devel/intel-graphics-compiler/intel-graphics-compiler-igc-1.0.9636/IGC/Compiler/CISACodeGen/ |
| H A D | helper.h | 422 unsigned simdWidth = 8; in GetThreadOccupancyPerSubslice() local
426 case SIMDMode::SIMD8: simdWidth = 8; break; in GetThreadOccupancyPerSubslice()
427 case SIMDMode::SIMD16: simdWidth = 16; break; in GetThreadOccupancyPerSubslice()
428 case SIMDMode::SIMD32: simdWidth = 32; break; in GetThreadOccupancyPerSubslice()
434 IGC_ASSERT(simdWidth); in GetThreadOccupancyPerSubslice()
435 const unsigned nThreadsPerTG = (threadGroupSize + simdWidth - 1) / simdWidth; in GetThreadOccupancyPerSubslice()
|
| /dports/graphics/krita/krita-4.4.8/libs/image/ |
| H A D | kis_brush_mask_applicators.h | 108 int simdWidth = width + alignOffset; in processVector() local
110 float *buffer = Vc::malloc<float, Vc::AlignOnCacheline>(simdWidth); in processVector()
116 …processor.template process<_impl>(buffer, simdWidth, y, m_d->cosa, m_d->sina, m_d->centerX, m_d->c… in processVector()
|
| /dports/lang/intel-compute-runtime/compute-runtime-21.52.22081/shared/source/kernel/ |
| H A D | implicit_args.h | 19 uint8_t simdWidth; member
|
| /dports/lang/intel-compute-runtime/compute-runtime-21.52.22081/opencl/source/built_ins/ |
| H A D | vme_dispatch_builder.h | 64 const size_t simdWidth = vmeKernel->getKernelInfo().getMaxSimdSize(); in buildDispatchInfos() local
73 Vec3<size_t> gws = {numThreadsX * simdWidth, 1, 1}; in buildDispatchInfos()
|
| /dports/lang/clover/mesa-21.3.6/src/gallium/drivers/swr/rasterizer/jitter/ |
| H A D | JitManager.cpp | 62 JitManager::JitManager(uint32_t simdWidth, const char* arch, const char* core) : in JitManager() argument
63 mContext(), mBuilder(mContext), mIsModuleFinalized(true), mJitNumber(0), mVWidth(simdWidth), in JitManager()
|
| /dports/graphics/libosmesa/mesa-21.3.6/src/gallium/drivers/swr/rasterizer/jitter/ |
| H A D | JitManager.cpp | 62 JitManager::JitManager(uint32_t simdWidth, const char* arch, const char* core) : in JitManager() argument
63 mContext(), mBuilder(mContext), mIsModuleFinalized(true), mJitNumber(0), mVWidth(simdWidth), in JitManager()
|