HomeSort by relevance Sort by last modified time
    Searched refs:Packet8f (Results 1 - 8 of 8) sorted by null

  /external/eigen/Eigen/src/Core/arch/AVX/
TypeCasting.h 39 template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
43 template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
PacketMath.h 31 typedef __m256 Packet8f;
40 const Packet8f p8f_##NAME = pset1<Packet8f>(X)
46 const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X))
56 typedef Packet8f type;
116 template<> struct unpacket_traits<Packet8f> { typedef float type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; };
120 template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float& from) { return _mm256_set1_ps(from); }
124 template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float* from) { return _mm256_broadcast_ss(from);
    [all...]
MathFunctions.h 32 inline Packet8f pshiftright(Packet8f v, int n)
48 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
49 psin<Packet8f>(const Packet8f& _x) {
50 Packet8f x = _x;
64 Packet8f z = pmul(x, p8f_one_over_pi);
65 Packet8f shift = _mm256_floor_ps(padd(z, p8f_one_over_four));
79 Packet8f ival_mask = _mm256_cmp_ps(z, p8f_one, _CMP_GT_OQ);
86 Packet8f z_minus_two = psub(z, p8f_two)
    [all...]
Complex.h 75 template<> EIGEN_STRONG_INLINE Packet4cf pload <Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet4cf(pload<Packet8f>(&numext::real_ref(*from))); }
76 template<> EIGEN_STRONG_INLINE Packet4cf ploadu<Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cf(ploadu<Packet8f>(&numext::real_ref(*from))); }
145 Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0));
146 Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0));
148 Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0));
149 Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0));
170 palign_impl<Offset*2,Packet8f>::run(first.v, second.v);
207 template<> struct conj_helper<Packet8f, Packet4cf, false,false>
209 EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const
212 EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) cons
    [all...]
  /external/tensorflow/tensorflow/core/kernels/
sparse_matmul_op.h 326 EIGEN_STRONG_INLINE Packet8f pinterleave4x64<Packet8f>(const Packet8f& from) {
344 EIGEN_STRONG_INLINE Packet8f pload4bf16<Packet8f>(const float* from) {
352 EIGEN_STRONG_INLINE Packet8f pload2bf16<Packet8f>(const float* from) {
381 EIGEN_DEVICE_FUNC inline Packet8f pexpand_bf16_l(const Packet8f& from) {
401 EIGEN_DEVICE_FUNC inline Packet8f pexpand_bf16_u(const Packet8f& from)
    [all...]
  /external/eigen/Eigen/src/Core/arch/CUDA/
TypeCasting.h 141 template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
154 template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
PacketMathHalf.h 784 EIGEN_STRONG_INLINE Packet8f half2float(const Packet8h& a) {
803 EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
829 Packet8f af = half2float(a);
830 Packet8f bf = half2float(b);
831 Packet8f rf = padd(af, bf);
836 Packet8f af = half2float(a);
837 Packet8f bf = half2float(b);
838 Packet8f rf = pmul(af, bf);
864 Packet8f af = half2float(a);
865 float reduced = predux<Packet8f>(af)
    [all...]
  /external/eigen/Eigen/src/Core/arch/AVX512/
PacketMath.h 51 typedef Packet8f half;
100 typedef Packet8f half;
464 Packet8f lane0 = _mm256_broadcast_ps((const __m128*)(const void*)from);
474 Packet8f lane1 = _mm256_broadcast_ps((const __m128*)(const void*)(from + 4));
881 Packet8f lane0 = _mm512_extractf32x8_ps(a, 0);
882 Packet8f lane1 = _mm512_extractf32x8_ps(a, 1);
883 Packet8f sum = padd(lane0, lane1);
884 Packet8f tmp0 = _mm256_hadd_ps(sum, _mm256_permute2f128_ps(a, a, 1));
908 EIGEN_STRONG_INLINE Packet8f predux_downto4<Packet16f>(const Packet16f& a) {
910 Packet8f lane0 = _mm512_extractf32x8_ps(a, 0)
    [all...]

Completed in 435 milliseconds