GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-12 03:39:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove unused packet op "preduxp".

Browse Source
This commit is contained in:
Rasmus Munk Larsen 2020-04-23 18:17:14 +00:00
parent 0aebe19aca
commit e80ec24357
16 changed files with 1 additions and 1009 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
Eigen/src/Core/GenericPacketMath.h
View File

@ -59,7 +59,6 @@ struct default_packet_traits
HasSetLinear = 1,
HasBlend = 0,
HasInsert = 0,
HasReduxp = 1,
HasDiv = 0,
HasSqrt = 0,
@ -485,10 +484,6 @@ template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* a
template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type pfirst(const Packet& a)
{ return a; }
/** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
preduxp(const Packet* vecs) { return vecs[0]; }
/** \internal \returns the sum of the elements of \a a*/
template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux(const Packet& a)
{ return a; }

23
Eigen/src/Core/arch/AVX/Complex.h
View File

@ -151,21 +151,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet4cf>(const Packe
Packet2cf(_mm256_extractf128_ps(a.v,1))));
}
template<> EIGEN_STRONG_INLINE Packet4cf preduxp<Packet4cf>(const Packet4cf* vecs)
{
Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0));
Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0));
t0 = _mm256_hadd_ps(t0,t1);
Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0));
Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0));
t2 = _mm256_hadd_ps(t2,t3);
t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4));
t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4));
return Packet4cf(_mm256_add_ps(t1,t3));
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet4cf>(const Packet4cf& a)
{
return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)),
@ -348,14 +333,6 @@ template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet2cd>(const Pack
Packet1cd(_mm256_extractf128_pd(a.v,1))));
}
template<> EIGEN_STRONG_INLINE Packet2cd preduxp<Packet2cd>(const Packet2cd* vecs)
{
Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4));
Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4));
return Packet2cd(_mm256_add_pd(t0,t1));
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet2cd>(const Packet2cd& a)
{
return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)),

57
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@ -627,49 +627,6 @@ template<> EIGEN_STRONG_INLINE Packet4d pldexp<Packet4d>(const Packet4d& a, cons
return pmul(a,_mm256_castsi256_pd(e));
}
// preduxp should be ok
// FIXME: why is this ok? why isn't the simply implementation working as expected?
template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs)
{
__m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]);
__m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]);
__m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]);
__m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]);
__m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
__m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
__m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
__m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
__m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
__m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
__m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
__m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
__m256 sum1 = _mm256_add_ps(perm1, hsum5);
__m256 sum2 = _mm256_add_ps(perm2, hsum6);
__m256 sum3 = _mm256_add_ps(perm3, hsum7);
__m256 sum4 = _mm256_add_ps(perm4, hsum8);
__m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
__m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
__m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
return final;
}
template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs)
{
Packet4d tmp0, tmp1;
tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]);
tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]);
tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
return _mm256_blend_pd(tmp0, tmp1, 0xC);
}
template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a)
{
return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1))));
@ -1105,20 +1062,6 @@ template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h&
return Eigen::half(reduced);
}
template<> EIGEN_STRONG_INLINE Packet8h preduxp<Packet8h>(const Packet8h* p) {
Packet8f pf[8];
pf[0] = half2float(p[0]);
pf[1] = half2float(p[1]);
pf[2] = half2float(p[2]);
pf[3] = half2float(p[3]);
pf[4] = half2float(p[4]);
pf[5] = half2float(p[5]);
pf[6] = half2float(p[6]);
pf[7] = half2float(p[7]);
Packet8f reduced = preduxp<Packet8f>(pf);
return float2half(reduced);
}
template<> EIGEN_STRONG_INLINE Packet8h preverse(const Packet8h& a)
{
__m128i m = _mm_setr_epi8(14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1);

2
Eigen/src/Core/arch/AVX512/Complex.h
View File

@ -42,7 +42,6 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits
HasMin = 0,
HasMax = 0,
HasSetLinear = 0,
HasReduxp = 0,
HasInsert = 1
};
};
@ -241,7 +240,6 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits
HasMin = 0,
HasMax = 0,
HasSetLinear = 0,
HasReduxp = 0
};
};

212
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@ -793,196 +793,6 @@ EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) {
OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 0), 2); \
OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 1), 3);
#endif
template<> EIGEN_STRONG_INLINE Packet16f preduxp<Packet16f>(const Packet16f*
vecs)
{
EIGEN_EXTRACT_8f_FROM_16f(vecs[0], vecs0);
EIGEN_EXTRACT_8f_FROM_16f(vecs[1], vecs1);
EIGEN_EXTRACT_8f_FROM_16f(vecs[2], vecs2);
EIGEN_EXTRACT_8f_FROM_16f(vecs[3], vecs3);
EIGEN_EXTRACT_8f_FROM_16f(vecs[4], vecs4);
EIGEN_EXTRACT_8f_FROM_16f(vecs[5], vecs5);
EIGEN_EXTRACT_8f_FROM_16f(vecs[6], vecs6);
EIGEN_EXTRACT_8f_FROM_16f(vecs[7], vecs7);
EIGEN_EXTRACT_8f_FROM_16f(vecs[8], vecs8);
EIGEN_EXTRACT_8f_FROM_16f(vecs[9], vecs9);
EIGEN_EXTRACT_8f_FROM_16f(vecs[10], vecs10);
EIGEN_EXTRACT_8f_FROM_16f(vecs[11], vecs11);
EIGEN_EXTRACT_8f_FROM_16f(vecs[12], vecs12);
EIGEN_EXTRACT_8f_FROM_16f(vecs[13], vecs13);
EIGEN_EXTRACT_8f_FROM_16f(vecs[14], vecs14);
EIGEN_EXTRACT_8f_FROM_16f(vecs[15], vecs15);
__m256 hsum1 = _mm256_hadd_ps(vecs0_0, vecs1_0);
__m256 hsum2 = _mm256_hadd_ps(vecs2_0, vecs3_0);
__m256 hsum3 = _mm256_hadd_ps(vecs4_0, vecs5_0);
__m256 hsum4 = _mm256_hadd_ps(vecs6_0, vecs7_0);
__m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
__m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
__m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
__m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
__m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
__m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
__m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
__m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
__m256 sum1 = _mm256_add_ps(perm1, hsum5);
__m256 sum2 = _mm256_add_ps(perm2, hsum6);
__m256 sum3 = _mm256_add_ps(perm3, hsum7);
__m256 sum4 = _mm256_add_ps(perm4, hsum8);
__m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
__m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
__m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
hsum1 = _mm256_hadd_ps(vecs0_1, vecs1_1);
hsum2 = _mm256_hadd_ps(vecs2_1, vecs3_1);
hsum3 = _mm256_hadd_ps(vecs4_1, vecs5_1);
hsum4 = _mm256_hadd_ps(vecs6_1, vecs7_1);
hsum5 = _mm256_hadd_ps(hsum1, hsum1);
hsum6 = _mm256_hadd_ps(hsum2, hsum2);
hsum7 = _mm256_hadd_ps(hsum3, hsum3);
hsum8 = _mm256_hadd_ps(hsum4, hsum4);
perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
sum1 = _mm256_add_ps(perm1, hsum5);
sum2 = _mm256_add_ps(perm2, hsum6);
sum3 = _mm256_add_ps(perm3, hsum7);
sum4 = _mm256_add_ps(perm4, hsum8);
blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
final = _mm256_add_ps(final, _mm256_blend_ps(blend1, blend2, 0xf0));
hsum1 = _mm256_hadd_ps(vecs8_0, vecs9_0);
hsum2 = _mm256_hadd_ps(vecs10_0, vecs11_0);
hsum3 = _mm256_hadd_ps(vecs12_0, vecs13_0);
hsum4 = _mm256_hadd_ps(vecs14_0, vecs15_0);
hsum5 = _mm256_hadd_ps(hsum1, hsum1);
hsum6 = _mm256_hadd_ps(hsum2, hsum2);
hsum7 = _mm256_hadd_ps(hsum3, hsum3);
hsum8 = _mm256_hadd_ps(hsum4, hsum4);
perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
sum1 = _mm256_add_ps(perm1, hsum5);
sum2 = _mm256_add_ps(perm2, hsum6);
sum3 = _mm256_add_ps(perm3, hsum7);
sum4 = _mm256_add_ps(perm4, hsum8);
blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
__m256 final_1 = _mm256_blend_ps(blend1, blend2, 0xf0);
hsum1 = _mm256_hadd_ps(vecs8_1, vecs9_1);
hsum2 = _mm256_hadd_ps(vecs10_1, vecs11_1);
hsum3 = _mm256_hadd_ps(vecs12_1, vecs13_1);
hsum4 = _mm256_hadd_ps(vecs14_1, vecs15_1);
hsum5 = _mm256_hadd_ps(hsum1, hsum1);
hsum6 = _mm256_hadd_ps(hsum2, hsum2);
hsum7 = _mm256_hadd_ps(hsum3, hsum3);
hsum8 = _mm256_hadd_ps(hsum4, hsum4);
perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
sum1 = _mm256_add_ps(perm1, hsum5);
sum2 = _mm256_add_ps(perm2, hsum6);
sum3 = _mm256_add_ps(perm3, hsum7);
sum4 = _mm256_add_ps(perm4, hsum8);
blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
final_1 = _mm256_add_ps(final_1, _mm256_blend_ps(blend1, blend2, 0xf0));
__m512 final_output;
EIGEN_INSERT_8f_INTO_16f(final_output, final, final_1);
return final_output;
}
template<> EIGEN_STRONG_INLINE Packet8d preduxp<Packet8d>(const Packet8d* vecs)
{
Packet4d vecs0_0 = _mm512_extractf64x4_pd(vecs[0], 0);
Packet4d vecs0_1 = _mm512_extractf64x4_pd(vecs[0], 1);
Packet4d vecs1_0 = _mm512_extractf64x4_pd(vecs[1], 0);
Packet4d vecs1_1 = _mm512_extractf64x4_pd(vecs[1], 1);
Packet4d vecs2_0 = _mm512_extractf64x4_pd(vecs[2], 0);
Packet4d vecs2_1 = _mm512_extractf64x4_pd(vecs[2], 1);
Packet4d vecs3_0 = _mm512_extractf64x4_pd(vecs[3], 0);
Packet4d vecs3_1 = _mm512_extractf64x4_pd(vecs[3], 1);
Packet4d vecs4_0 = _mm512_extractf64x4_pd(vecs[4], 0);
Packet4d vecs4_1 = _mm512_extractf64x4_pd(vecs[4], 1);
Packet4d vecs5_0 = _mm512_extractf64x4_pd(vecs[5], 0);
Packet4d vecs5_1 = _mm512_extractf64x4_pd(vecs[5], 1);
Packet4d vecs6_0 = _mm512_extractf64x4_pd(vecs[6], 0);
Packet4d vecs6_1 = _mm512_extractf64x4_pd(vecs[6], 1);
Packet4d vecs7_0 = _mm512_extractf64x4_pd(vecs[7], 0);
Packet4d vecs7_1 = _mm512_extractf64x4_pd(vecs[7], 1);
Packet4d tmp0, tmp1;
tmp0 = _mm256_hadd_pd(vecs0_0, vecs1_0);
tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
tmp1 = _mm256_hadd_pd(vecs2_0, vecs3_0);
tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
__m256d final_0 = _mm256_blend_pd(tmp0, tmp1, 0xC);
tmp0 = _mm256_hadd_pd(vecs0_1, vecs1_1);
tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
tmp1 = _mm256_hadd_pd(vecs2_1, vecs3_1);
tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
final_0 = _mm256_add_pd(final_0, _mm256_blend_pd(tmp0, tmp1, 0xC));
tmp0 = _mm256_hadd_pd(vecs4_0, vecs5_0);
tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
tmp1 = _mm256_hadd_pd(vecs6_0, vecs7_0);
tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
__m256d final_1 = _mm256_blend_pd(tmp0, tmp1, 0xC);
tmp0 = _mm256_hadd_pd(vecs4_1, vecs5_1);
tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
tmp1 = _mm256_hadd_pd(vecs6_1, vecs7_1);
tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
final_1 = _mm256_add_pd(final_1, _mm256_blend_pd(tmp0, tmp1, 0xC));
__m512d final_output = _mm512_castpd256_pd512(final_0);
return _mm512_insertf64x4(final_output, final_1, 1);
}
template <>
EIGEN_STRONG_INLINE float predux<Packet16f>(const Packet16f& a) {
@ -1660,28 +1470,6 @@ template<> EIGEN_STRONG_INLINE half predux_mul<Packet16h>(const Packet16h& from)
return half(predux_mul(from_float));
}
template<> EIGEN_STRONG_INLINE Packet16h preduxp<Packet16h>(const Packet16h* p) {
Packet16f pf[16];
pf[0] = half2float(p[0]);
pf[1] = half2float(p[1]);
pf[2] = half2float(p[2]);
pf[3] = half2float(p[3]);
pf[4] = half2float(p[4]);
pf[5] = half2float(p[5]);
pf[6] = half2float(p[6]);
pf[7] = half2float(p[7]);
pf[8] = half2float(p[8]);
pf[9] = half2float(p[9]);
pf[10] = half2float(p[10]);
pf[11] = half2float(p[11]);
pf[12] = half2float(p[12]);
pf[13] = half2float(p[13]);
pf[14] = half2float(p[14]);
pf[15] = half2float(p[15]);
Packet16f reduced = preduxp<Packet16f>(pf);
return float2half(reduced);
}
template<> EIGEN_STRONG_INLINE Packet16h preverse(const Packet16h& a)
{
__m128i m = _mm_setr_epi8(14,15,12,13,10,11,8,9,6,7,4,5,2,3,0,1);

17
Eigen/src/Core/arch/AltiVec/Complex.h
View File

@ -149,22 +149,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe
return pfirst<Packet2cf>(Packet2cf(b));
}
template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
{
Packet4f b1, b2;
#ifdef _BIG_ENDIAN
b1 = vec_sld(vecs[0].v, vecs[1].v, 8);
b2 = vec_sld(vecs[1].v, vecs[0].v, 8);
#else
b1 = vec_sld(vecs[1].v, vecs[0].v, 8);
b2 = vec_sld(vecs[0].v, vecs[1].v, 8);
#endif
b2 = vec_sld(b2, b2, 8);
b2 = padd<Packet4f>(b1, b2);
return Packet2cf(b2);
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
{
Packet4f b;
@ -359,7 +343,6 @@ template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Pac
template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; }
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }

142
Eigen/src/Core/arch/AltiVec/PacketMath.h
View File

@ -929,106 +929,6 @@ template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
return pfirst(sum);
}
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
Packet4f v[4], sum[4];
// It's easier and faster to transpose then add as columns
// Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
// Do the transpose, first set of moves
v[0] = vec_mergeh(vecs[0], vecs[2]);
v[1] = vec_mergel(vecs[0], vecs[2]);
v[2] = vec_mergeh(vecs[1], vecs[3]);
v[3] = vec_mergel(vecs[1], vecs[3]);
// Get the resulting vectors
sum[0] = vec_mergeh(v[0], v[2]);
sum[1] = vec_mergel(v[0], v[2]);
sum[2] = vec_mergeh(v[1], v[3]);
sum[3] = vec_mergel(v[1], v[3]);
// Now do the summation:
// Lines 0+1
sum[0] = sum[0] + sum[1];
// Lines 2+3
sum[1] = sum[2] + sum[3];
// Add the results
sum[0] = sum[0] + sum[1];
return sum[0];
}
template<> EIGEN_STRONG_INLINE Packet8s preduxp<Packet8s>(const Packet8s* vecs)
{
Packet8s step1[8], step2[8], step3[8];
step1[0] = vec_mergeh(vecs[0], vecs[4]);
step1[1] = vec_mergel(vecs[0], vecs[4]);
step1[2] = vec_mergeh(vecs[1], vecs[5]);
step1[3] = vec_mergel(vecs[1], vecs[5]);
step1[4] = vec_mergeh(vecs[2], vecs[6]);
step1[5] = vec_mergel(vecs[2], vecs[6]);
step1[6] = vec_mergeh(vecs[3], vecs[7]);
step1[7] = vec_mergel(vecs[3], vecs[7]);
step2[0] = vec_mergeh(step1[0], step1[4]);
step2[1] = vec_mergel(step1[0], step1[4]);
step2[2] = vec_mergeh(step1[1], step1[5]);
step2[3] = vec_mergel(step1[1], step1[5]);
step2[4] = vec_mergeh(step1[2], step1[6]);
step2[5] = vec_mergel(step1[2], step1[6]);
step2[6] = vec_mergeh(step1[3], step1[7]);
step2[7] = vec_mergel(step1[3], step1[7]);
step3[0] = vec_mergeh(step2[0], step2[4]);
step3[1] = vec_mergel(step2[0], step2[4]);
step3[2] = vec_mergeh(step2[1], step2[5]);
step3[3] = vec_mergel(step2[1], step2[5]);
step3[4] = vec_mergeh(step2[2], step2[6]);
step3[5] = vec_mergel(step2[2], step2[6]);
step3[6] = vec_mergeh(step2[3], step2[7]);
step3[7] = vec_mergel(step2[3], step2[7]);
step3[0] += step3[1] + step3[2] + step3[3] + step3[4] + step3[5] + step3[6] + step3[7];
return step3[0];
}
template<> EIGEN_STRONG_INLINE Packet8us preduxp<Packet8us>(const Packet8us* vecs)
{
Packet8us step1[8], step2[8], step3[8];
step1[0] = vec_mergeh(vecs[0], vecs[4]);
step1[1] = vec_mergel(vecs[0], vecs[4]);
step1[2] = vec_mergeh(vecs[1], vecs[5]);
step1[3] = vec_mergel(vecs[1], vecs[5]);
step1[4] = vec_mergeh(vecs[2], vecs[6]);
step1[5] = vec_mergel(vecs[2], vecs[6]);
step1[6] = vec_mergeh(vecs[3], vecs[7]);
step1[7] = vec_mergel(vecs[3], vecs[7]);
step2[0] = vec_mergeh(step1[0], step1[4]);
step2[1] = vec_mergel(step1[0], step1[4]);
step2[2] = vec_mergeh(step1[1], step1[5]);
step2[3] = vec_mergel(step1[1], step1[5]);
step2[4] = vec_mergeh(step1[2], step1[6]);
step2[5] = vec_mergel(step1[2], step1[6]);
step2[6] = vec_mergeh(step1[3], step1[7]);
step2[7] = vec_mergel(step1[3], step1[7]);
step3[0] = vec_mergeh(step2[0], step2[4]);
step3[1] = vec_mergel(step2[0], step2[4]);
step3[2] = vec_mergeh(step2[1], step2[5]);
step3[3] = vec_mergel(step2[1], step2[5]);
step3[4] = vec_mergeh(step2[2], step2[6]);
step3[5] = vec_mergel(step2[2], step2[6]);
step3[6] = vec_mergeh(step2[3], step2[7]);
step3[7] = vec_mergel(step2[3], step2[7]);
step3[0] += step3[1] + step3[2] + step3[3] + step3[4] + step3[5] + step3[6] + step3[7];
return step3[0];
}
template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
{
Packet4i sum;
@ -1074,34 +974,6 @@ template<> EIGEN_STRONG_INLINE unsigned short int predux<Packet8us>(const Packet
return static_cast<unsigned short int>(predux(first_half) + predux(second_half));
}
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
Packet4i v[4], sum[4];
// It's easier and faster to transpose then add as columns
// Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
// Do the transpose, first set of moves
v[0] = vec_mergeh(vecs[0], vecs[2]);
v[1] = vec_mergel(vecs[0], vecs[2]);
v[2] = vec_mergeh(vecs[1], vecs[3]);
v[3] = vec_mergel(vecs[1], vecs[3]);
// Get the resulting vectors
sum[0] = vec_mergeh(v[0], v[2]);
sum[1] = vec_mergel(v[0], v[2]);
sum[2] = vec_mergeh(v[1], v[3]);
sum[3] = vec_mergel(v[1], v[3]);
// Now do the summation:
// Lines 0+1
sum[0] = sum[0] + sum[1];
// Lines 2+3
sum[1] = sum[2] + sum[3];
// Add the results
sum[0] = sum[0] + sum[1];
return sum[0];
}
// Other reduction functions:
// mul
template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
@ -1835,20 +1707,6 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
return pfirst<Packet2d>(sum);
}
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
Packet2d v[2], sum;
v[0] = vecs[0] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[0]), reinterpret_cast<Packet4f>(vecs[0]), 8));
v[1] = vecs[1] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[1]), reinterpret_cast<Packet4f>(vecs[1]), 8));
#ifdef _BIG_ENDIAN
sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[0]), reinterpret_cast<Packet4f>(v[1]), 8));
#else
sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[1]), reinterpret_cast<Packet4f>(v[0]), 8));
#endif
return sum;
}
// Other reduction functions:
// mul
template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)

21
Eigen/src/Core/arch/MSA/Complex.h
View File

@ -297,20 +297,6 @@ EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) {
return std::complex<float>(value[0], value[1]);
}
template <>
EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) {
EIGEN_MSA_DEBUG;
Packet4f sum1, sum2, sum;
// Add the first two 64-bit float32x2_t of vecs[0]
sum1 = (Packet4f)__builtin_msa_ilvr_d((v2i64)vecs[1].v, (v2i64)vecs[0].v);
sum2 = (Packet4f)__builtin_msa_ilvl_d((v2i64)vecs[1].v, (v2i64)vecs[0].v);
sum = padd(sum1, sum2);
return Packet2cf(sum);
}
template <>
EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) {
EIGEN_MSA_DEBUG;
@ -660,13 +646,6 @@ EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) {
return pfirst(a);
}
template <>
EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) {
EIGEN_MSA_DEBUG;
return vecs[0];
}
template <>
EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) {
EIGEN_MSA_DEBUG;

49
Eigen/src/Core/arch/MSA/PacketMath.h
View File

@ -575,45 +575,6 @@ EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) {
return s[0];
}
template <>
EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) {
EIGEN_MSA_DEBUG;
v4i32 tmp1, tmp2, tmp3, tmp4;
Packet4f sum;
tmp1 = __builtin_msa_ilvr_w((v4i32)vecs[1], (v4i32)vecs[0]);
tmp2 = __builtin_msa_ilvr_w((v4i32)vecs[3], (v4i32)vecs[2]);
tmp3 = __builtin_msa_ilvl_w((v4i32)vecs[1], (v4i32)vecs[0]);
tmp4 = __builtin_msa_ilvl_w((v4i32)vecs[3], (v4i32)vecs[2]);
sum = (Packet4f)__builtin_msa_ilvr_d((v2i64)tmp2, (v2i64)tmp1);
sum = padd(sum, (Packet4f)__builtin_msa_ilvod_d((v2i64)tmp2, (v2i64)tmp1));
sum = padd(sum, (Packet4f)__builtin_msa_ilvr_d((v2i64)tmp4, (v2i64)tmp3));
sum = padd(sum, (Packet4f)__builtin_msa_ilvod_d((v2i64)tmp4, (v2i64)tmp3));
return sum;
}
template <>
EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) {
EIGEN_MSA_DEBUG;
v4i32 tmp1, tmp2, tmp3, tmp4;
Packet4i sum;
tmp1 = __builtin_msa_ilvr_w((v4i32)vecs[1], (v4i32)vecs[0]);
tmp2 = __builtin_msa_ilvr_w((v4i32)vecs[3], (v4i32)vecs[2]);
tmp3 = __builtin_msa_ilvl_w((v4i32)vecs[1], (v4i32)vecs[0]);
tmp4 = __builtin_msa_ilvl_w((v4i32)vecs[3], (v4i32)vecs[2]);
sum = (Packet4i)__builtin_msa_ilvr_d((v2i64)tmp2, (v2i64)tmp1);
sum = padd(sum, (Packet4i)__builtin_msa_ilvod_d((v2i64)tmp2, (v2i64)tmp1));
sum = padd(sum, (Packet4i)__builtin_msa_ilvr_d((v2i64)tmp4, (v2i64)tmp3));
sum = padd(sum, (Packet4i)__builtin_msa_ilvod_d((v2i64)tmp4, (v2i64)tmp3));
return sum;
}
template <>
EIGEN_STRONG_INLINE int32_t predux<Packet4i>(const Packet4i& a) {
@ -1148,16 +1109,6 @@ EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) {
return s[0];
}
template <>
EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) {
EIGEN_MSA_DEBUG;
Packet2d v0 = (Packet2d)__builtin_msa_ilvev_d((v2i64)vecs[1], (v2i64)vecs[0]);
Packet2d v1 = (Packet2d)__builtin_msa_ilvod_d((v2i64)vecs[1], (v2i64)vecs[0]);
return padd(v0, v1);
}
// Other reduction functions:
// mul
template <>

10
Eigen/src/Core/arch/NEON/Complex.h
View File

@ -307,14 +307,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe
return s;
}
template<> EIGEN_STRONG_INLINE Packet1cf preduxp<Packet1cf>(const Packet1cf* vecs) { return vecs[0]; }
template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
{
const Packet4f sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v));
const Packet4f sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v));
return Packet2cf(vaddq_f32(sum1, sum2));
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet1cf>(const Packet1cf& a)
{
std::complex<float> s;
@ -608,8 +600,6 @@ template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a
template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; }
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
template<int Offset>

289
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@ -137,7 +137,6 @@ struct packet_traits<float> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1,
HasDiv = 1,
HasFloor = 1,
@ -180,7 +179,6 @@ struct packet_traits<int8_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1
};
};
@ -212,7 +210,6 @@ struct packet_traits<uint8_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1,
HasSqrt = 1
};
@ -246,7 +243,6 @@ struct packet_traits<int16_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1
};
};
@ -278,7 +274,6 @@ struct packet_traits<uint16_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1,
HasSqrt = 1
};
@ -312,7 +307,6 @@ struct packet_traits<int32_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1
};
};
@ -344,7 +338,6 @@ struct packet_traits<uint32_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1,
HasSqrt = 1
};
@ -379,7 +372,6 @@ struct packet_traits<int64_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1
};
};
@ -412,7 +404,6 @@ struct packet_traits<uint64_t> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1
};
};
@ -2422,281 +2413,6 @@ template<> EIGEN_STRONG_INLINE int64_t predux<Packet2l>(const Packet2l& a)
template<> EIGEN_STRONG_INLINE uint64_t predux<Packet2ul>(const Packet2ul& a)
{ return vgetq_lane_u64(a, 0) + vgetq_lane_u64(a, 1); }
template<> EIGEN_STRONG_INLINE Packet2f preduxp<Packet2f>(const Packet2f* vecs)
{
const float32x2x2_t vtrn = vzip_f32(vecs[0], vecs[1]);
return vadd_f32(vtrn.val[0], vtrn.val[1]);
}
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
const float32x4x2_t vtrn1 = vzipq_f32(vecs[0], vecs[2]);
const float32x4x2_t vtrn2 = vzipq_f32(vecs[1], vecs[3]);
const float32x4x2_t res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]);
const float32x4x2_t res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]);
return vaddq_f32(vaddq_f32(res1.val[0], res1.val[1]), vaddq_f32(res2.val[0], res2.val[1]));
}
template<> EIGEN_STRONG_INLINE Packet4c preduxp<Packet4c>(const Packet4c* vecs)
{
const int8x8x2_t zip8 = vzip_s8(
vreinterpret_s8_s32(vset_lane_s32(vecs[2], vdup_n_s32(vecs[0]), 1)),
vreinterpret_s8_s32(vset_lane_s32(vecs[3], vdup_n_s32(vecs[1]), 1)));
const uint16x4x2_t zip16 = vzip_u16(
vreinterpret_u16_s8(zip8.val[0]),
vreinterpret_u16_s8(zip8.val[1]));
const int8x8_t sum = vadd_s8(
vreinterpret_s8_u16(zip16.val[0]),
vreinterpret_s8_u16(zip16.val[1]));
return vget_lane_s32(vreinterpret_s32_s8(vadd_s8(sum,
vreinterpret_s8_s32(vrev64_s32(vreinterpret_s32_s8(sum))))), 0);
}
template<> EIGEN_STRONG_INLINE Packet8c preduxp<Packet8c>(const Packet8c* vecs)
{
int8x8_t sum[4];
EIGEN_UNROLL_LOOP
for (int i = 0; i != 4; i++)
{
const int8x8x2_t z = vzip_s8(vecs[i*2], vecs[i*2+1]);
sum[i] = vadd_s8(z.val[0], z.val[1]);
}
EIGEN_UNROLL_LOOP
for (int i = 0; i != 2; i++)
{
const uint16x4x2_t z = vzip_u16(vreinterpret_u16_s8(sum[i*2]), vreinterpret_u16_s8(sum[i*2+1]));
sum[i] = vadd_s8(vreinterpret_s8_u16(z.val[0]), vreinterpret_s8_u16(z.val[1]));
}
const uint32x2x2_t z = vzip_u32(vreinterpret_u32_s8(sum[0]), vreinterpret_u32_s8(sum[1]));
return vadd_s8(vreinterpret_s8_u32(z.val[0]), vreinterpret_s8_u32(z.val[1]));
}
template<> EIGEN_STRONG_INLINE Packet16c preduxp<Packet16c>(const Packet16c* vecs)
{
int8x16_t sum[8];
EIGEN_UNROLL_LOOP
for (int i = 0; i != 8; i++)
{
const int8x16x2_t z = vzipq_s8(vecs[i*2], vecs[i*2+1]);
sum[i] = vaddq_s8(z.val[0], z.val[1]);
}
EIGEN_UNROLL_LOOP
for (int i = 0; i != 4; i++)
{
const uint16x8x2_t z = vzipq_u16(vreinterpretq_u16_s8(sum[i*2]), vreinterpretq_u16_s8(sum[i*2+1]));
sum[i] = vaddq_s8(vreinterpretq_s8_u16(z.val[0]), vreinterpretq_s8_u16(z.val[1]));
}
EIGEN_UNROLL_LOOP
for (int i = 0; i != 2; i++)
{
const uint32x4x2_t z = vzipq_u32(vreinterpretq_u32_s8(sum[i*2]), vreinterpretq_u32_s8(sum[i*2+1]));
sum[i] = vaddq_s8(vreinterpretq_s8_u32(z.val[0]), vreinterpretq_s8_u32(z.val[1]));
}
return vcombine_s8(
vadd_s8(vget_low_s8(sum[0]), vget_high_s8(sum[0])),
vadd_s8(vget_low_s8(sum[1]), vget_high_s8(sum[1])));
}
template<> EIGEN_STRONG_INLINE Packet4uc preduxp<Packet4uc>(const Packet4uc* vecs)
{
const uint8x8x2_t zip8 = vzip_u8(
vreinterpret_u8_u32(vset_lane_u32(vecs[2], vdup_n_u32(vecs[0]), 1)),
vreinterpret_u8_u32(vset_lane_u32(vecs[3], vdup_n_u32(vecs[1]), 1)));
const uint16x4x2_t zip16 = vzip_u16(
vreinterpret_u16_u8(zip8.val[0]),
vreinterpret_u16_u8(zip8.val[1]));
const uint8x8_t sum = vadd_u8(
vreinterpret_u8_u16(zip16.val[0]),
vreinterpret_u8_u16(zip16.val[1]));
return vget_lane_u32(vreinterpret_u32_u8(vadd_u8(sum,
vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(sum))))), 0);
}
template<> EIGEN_STRONG_INLINE Packet8uc preduxp<Packet8uc>(const Packet8uc* vecs)
{
uint8x8_t sum[4];
EIGEN_UNROLL_LOOP
for (int i = 0; i != 4; i++)
{
const uint8x8x2_t z = vzip_u8(vecs[i*2], vecs[i*2+1]);
sum[i] = vadd_u8(z.val[0], z.val[1]);
}
EIGEN_UNROLL_LOOP
for (int i = 0; i != 2; i++)
{
const uint16x4x2_t z = vzip_u16(vreinterpret_u16_u8(sum[i*2]), vreinterpret_u16_u8(sum[i*2+1]));
sum[i] = vadd_u8(vreinterpret_u8_u16(z.val[0]), vreinterpret_u8_u16(z.val[1]));
}
const uint32x2x2_t z = vzip_u32(vreinterpret_u32_u8(sum[0]), vreinterpret_u32_u8(sum[1]));
return vadd_u8(vreinterpret_u8_u32(z.val[0]), vreinterpret_u8_u32(z.val[1]));
}
template<> EIGEN_STRONG_INLINE Packet16uc preduxp<Packet16uc>(const Packet16uc* vecs)
{
uint8x16_t sum[8];
EIGEN_UNROLL_LOOP
for (int i = 0; i != 8; i++)
{
const uint8x16x2_t z = vzipq_u8(vecs[i*2], vecs[i*2+1]);
sum[i] = vaddq_u8(z.val[0], z.val[1]);
}
EIGEN_UNROLL_LOOP
for (int i = 0; i != 4; i++)
{
const uint16x8x2_t z = vzipq_u16(vreinterpretq_u16_u8(sum[i*2]), vreinterpretq_u16_u8(sum[i*2+1]));
sum[i] = vaddq_u8(vreinterpretq_u8_u16(z.val[0]), vreinterpretq_u8_u16(z.val[1]));
}
EIGEN_UNROLL_LOOP
for (int i = 0; i != 2; i++)
{
const uint32x4x2_t z = vzipq_u32(vreinterpretq_u32_u8(sum[i*2]), vreinterpretq_u32_u8(sum[i*2+1]));
sum[i] = vaddq_u8(vreinterpretq_u8_u32(z.val[0]), vreinterpretq_u8_u32(z.val[1]));
}
return vcombine_u8(
vadd_u8(vget_low_u8(sum[0]), vget_high_u8(sum[0])),
vadd_u8(vget_low_u8(sum[1]), vget_high_u8(sum[1])));
}
template<> EIGEN_STRONG_INLINE Packet4s preduxp<Packet4s>(const Packet4s* vecs)
{
int16x4x2_t zip16;
int32x2x2_t zip32;
int16x4_t sum1, sum2;
zip16 = vzip_s16(vecs[0], vecs[1]);
sum1 = vadd_s16(zip16.val[0], zip16.val[1]);
zip16 = vzip_s16(vecs[2], vecs[3]);
sum2 = vadd_s16(zip16.val[0], zip16.val[1]);
zip32 = vzip_s32(vreinterpret_s32_s16(sum1), vreinterpret_s32_s16(sum2));
return vadd_s16(vreinterpret_s16_s32(zip32.val[0]), vreinterpret_s16_s32(zip32.val[1]));
}
template<> EIGEN_STRONG_INLINE Packet8s preduxp<Packet8s>(const Packet8s* vecs)
{
int16x8x2_t zip16;
int32x4x2_t zip32;
int16x8_t sum1, sum2, sum3, sum4;
zip16 = vzipq_s16(vecs[0], vecs[1]);
sum1 = vaddq_s16(zip16.val[0], zip16.val[1]);
zip16 = vzipq_s16(vecs[2], vecs[3]);
sum2 = vaddq_s16(zip16.val[0], zip16.val[1]);
zip16 = vzipq_s16(vecs[4], vecs[5]);
sum3 = vaddq_s16(zip16.val[0], zip16.val[1]);
zip16 = vzipq_s16(vecs[6], vecs[7]);
sum4 = vaddq_s16(zip16.val[0], zip16.val[1]);
zip32 = vzipq_s32(vreinterpretq_s32_s16(sum1), vreinterpretq_s32_s16(sum2));
sum1 = vaddq_s16(vreinterpretq_s16_s32(zip32.val[0]), vreinterpretq_s16_s32(zip32.val[1]));
zip32 = vzipq_s32(vreinterpretq_s32_s16(sum3), vreinterpretq_s32_s16(sum4));
sum2 = vaddq_s16(vreinterpretq_s16_s32(zip32.val[0]), vreinterpretq_s16_s32(zip32.val[1]));
return vcombine_s16(
vadd_s16(vget_low_s16(sum1), vget_high_s16(sum1)),
vadd_s16(vget_low_s16(sum2), vget_high_s16(sum2)));
}
template<> EIGEN_STRONG_INLINE Packet4us preduxp<Packet4us>(const Packet4us* vecs)
{
uint16x4x2_t zip16;
uint32x2x2_t zip32;
uint16x4_t sum1, sum2;
zip16 = vzip_u16(vecs[0], vecs[1]);
sum1 = vadd_u16(zip16.val[0], zip16.val[1]);
zip16 = vzip_u16(vecs[2], vecs[3]);
sum2 = vadd_u16(zip16.val[0], zip16.val[1]);
zip32 = vzip_u32(vreinterpret_u32_u16(sum1), vreinterpret_u32_u16(sum2));
return vadd_u16(vreinterpret_u16_u32(zip32.val[0]), vreinterpret_u16_u32(zip32.val[1]));
}
template<> EIGEN_STRONG_INLINE Packet8us preduxp<Packet8us>(const Packet8us* vecs)
{
uint16x8x2_t zip16;
uint32x4x2_t zip32;
uint16x8_t sum1, sum2, sum3, sum4;
zip16 = vzipq_u16(vecs[0], vecs[1]);
sum1 = vaddq_u16(zip16.val[0], zip16.val[1]);
zip16 = vzipq_u16(vecs[2], vecs[3]);
sum2 = vaddq_u16(zip16.val[0], zip16.val[1]);
zip16 = vzipq_u16(vecs[4], vecs[5]);
sum3 = vaddq_u16(zip16.val[0], zip16.val[1]);
zip16 = vzipq_u16(vecs[6], vecs[7]);
sum4 = vaddq_u16(zip16.val[0], zip16.val[1]);
zip32 = vzipq_u32(vreinterpretq_u32_u16(sum1), vreinterpretq_u32_u16(sum2));
sum1 = vaddq_u16(vreinterpretq_u16_u32(zip32.val[0]), vreinterpretq_u16_u32(zip32.val[1]));
zip32 = vzipq_u32(vreinterpretq_u32_u16(sum3), vreinterpretq_u32_u16(sum4));
sum2 = vaddq_u16(vreinterpretq_u16_u32(zip32.val[0]), vreinterpretq_u16_u32(zip32.val[1]));
return vcombine_u16(
vadd_u16(vget_low_u16(sum1), vget_high_u16(sum1)),
vadd_u16(vget_low_u16(sum2), vget_high_u16(sum2)));
}
template<> EIGEN_STRONG_INLINE Packet2i preduxp<Packet2i>(const Packet2i* vecs)
{
const int32x2x2_t vtrn = vzip_s32(vecs[0], vecs[1]);
return vadd_s32(vtrn.val[0], vtrn.val[1]);
}
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
const int32x4x2_t vtrn1 = vzipq_s32(vecs[0], vecs[2]);
const int32x4x2_t vtrn2 = vzipq_s32(vecs[1], vecs[3]);
const int32x4x2_t res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]);
const int32x4x2_t res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]);
return vaddq_s32(vaddq_s32(res1.val[0], res1.val[1]), vaddq_s32(res2.val[0], res2.val[1]));
}
template<> EIGEN_STRONG_INLINE Packet2ui preduxp<Packet2ui>(const Packet2ui* vecs)
{
const uint32x2x2_t vtrn = vzip_u32(vecs[0], vecs[1]);
return vadd_u32(vtrn.val[0], vtrn.val[1]);
}
template<> EIGEN_STRONG_INLINE Packet4ui preduxp<Packet4ui>(const Packet4ui* vecs)
{
uint32x4x2_t vtrn1, vtrn2, res1, res2;
Packet4ui sum1, sum2, sum;
// NEON zip performs interleaving of the supplied vectors.
// We perform two interleaves in a row to acquire the transposed vector
vtrn1 = vzipq_u32(vecs[0], vecs[2]);
vtrn2 = vzipq_u32(vecs[1], vecs[3]);
res1 = vzipq_u32(vtrn1.val[0], vtrn2.val[0]);
res2 = vzipq_u32(vtrn1.val[1], vtrn2.val[1]);
// Do the addition of the resulting vectors
sum1 = vaddq_u32(res1.val[0], res1.val[1]);
sum2 = vaddq_u32(res2.val[0], res2.val[1]);
sum = vaddq_u32(sum1, sum2);
return sum;
}
template<> EIGEN_STRONG_INLINE Packet2l preduxp<Packet2l>(const Packet2l* vecs)
{
return vsetq_lane_s64(
vget_lane_s64(vget_low_s64(vecs[0]), 0) +
vget_lane_s64(vget_high_s64(vecs[0]), 0),
vdupq_n_s64(
vget_lane_s64(vget_low_s64(vecs[1]), 0) +
vget_lane_s64(vget_high_s64(vecs[1]), 0)),
0);
}
template<> EIGEN_STRONG_INLINE Packet2ul preduxp<Packet2ul>(const Packet2ul* vecs)
{
return vsetq_lane_u64(
vget_lane_u64(vget_low_u64(vecs[0]), 0) +
vget_lane_u64(vget_high_u64(vecs[0]), 0),
vdupq_n_u64(
vget_lane_u64(vget_low_u64(vecs[1]), 0) +
vget_lane_u64(vget_high_u64(vecs[1]), 0)),
0);
}
template<> EIGEN_DEVICE_FUNC inline Packet4c predux_half_dowto4(const Packet8c& a)
{
return vget_lane_s32(vreinterpret_s32_s8(vadd_s8(a,
@ -3687,7 +3403,6 @@ template<> struct packet_traits<double> : default_packet_traits
HasSetLinear = 0,
HasBlend = 0,
HasInsert = 1,
HasReduxp = 1,
HasDiv = 1,
HasFloor = 1,
@ -3830,10 +3545,6 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
{ return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0); }
#endif
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
return vaddq_f64(vzip1q_f64(vecs[0], vecs[1]), vzip2q_f64(vecs[0], vecs[1]));
}
// Other reduction functions:
// mul
#if EIGEN_COMP_CLANG && defined(__apple_build_version__)

10
Eigen/src/Core/arch/SSE/Complex.h
View File

@ -156,11 +156,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe
return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
}
template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
{
return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)));
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
{
return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v))));
@ -346,11 +341,6 @@ template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Pack
return pfirst(a);
}
template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
{
return vecs[0];
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
{
return pfirst(a);

53
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@ -179,8 +179,7 @@ template<> struct packet_traits<bool> : default_packet_traits
HasAbs2 = 0,
HasMin = 0,
HasMax = 0,
HasConj = 0,
HasReduxp = 0
HasConj = 0
};
};
@ -733,38 +732,6 @@ EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
}
#ifdef EIGEN_VECTORIZE_SSE3
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
}
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
return _mm_hadd_pd(vecs[0], vecs[1]);
}
#else
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
Packet4f tmp0, tmp1, tmp2;
tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]);
tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]);
tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]);
tmp0 = _mm_add_ps(tmp0, tmp1);
tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]);
tmp1 = _mm_add_ps(tmp1, tmp2);
tmp2 = _mm_movehl_ps(tmp1, tmp0);
tmp0 = _mm_movelh_ps(tmp0, tmp1);
return _mm_add_ps(tmp0, tmp2);
}
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1]));
}
#endif // SSE3
template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
{
// Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
@ -790,10 +757,6 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
}
#ifdef EIGEN_VECTORIZE_SSSE3
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
}
template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
{
Packet4i tmp0 = _mm_hadd_epi32(a,a);
@ -805,20 +768,6 @@ template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
}
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
Packet4i tmp0, tmp1, tmp2;
tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
tmp0 = _mm_add_epi32(tmp0, tmp1);
tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
tmp1 = _mm_add_epi32(tmp1, tmp2);
tmp2 = _mm_unpacklo_epi64(tmp0, tmp1);
tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
return _mm_add_epi32(tmp0, tmp2);
}
#endif
// Other reduction functions:

25
Eigen/src/Core/arch/ZVector/Complex.h
View File

@ -156,10 +156,6 @@ template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Pack
{
return pfirst(a);
}
template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
{
return vecs[0];
}
template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
{
return pfirst(a);
@ -327,16 +323,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe
return res;
}
template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
{
PacketBlock<Packet2cf,2> transpose;
transpose.packet[0] = vecs[0];
transpose.packet[1] = vecs[1];
ptranspose(transpose);
return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
{
std::complex<float> res;
@ -461,17 +447,6 @@ template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packe
return pfirst<Packet2cf>(Packet2cf(b));
}
template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
{
Packet4f b1, b2;
b1 = vec_sld(vecs[0].v, vecs[1].v, 8);
b2 = vec_sld(vecs[1].v, vecs[0].v, 8);
b2 = vec_sld(b2, b2, 8);
b2 = padd<Packet4f>(b1, b2);
return Packet2cf(b2);
}
template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
{
Packet4f b;

82
Eigen/src/Core/arch/ZVector/PacketMath.h
View File

@ -530,45 +530,6 @@ template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
return pfirst(sum);
}
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
Packet4i v[4], sum[4];
// It's easier and faster to transpose then add as columns
// Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
// Do the transpose, first set of moves
v[0] = vec_mergeh(vecs[0], vecs[2]);
v[1] = vec_mergel(vecs[0], vecs[2]);
v[2] = vec_mergeh(vecs[1], vecs[3]);
v[3] = vec_mergel(vecs[1], vecs[3]);
// Get the resulting vectors
sum[0] = vec_mergeh(v[0], v[2]);
sum[1] = vec_mergel(v[0], v[2]);
sum[2] = vec_mergeh(v[1], v[3]);
sum[3] = vec_mergel(v[1], v[3]);
// Now do the summation:
// Lines 0+1
sum[0] = padd<Packet4i>(sum[0], sum[1]);
// Lines 2+3
sum[1] = padd<Packet4i>(sum[2], sum[3]);
// Add the results
sum[0] = padd<Packet4i>(sum[0], sum[1]);
return sum[0];
}
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
Packet2d v[2], sum;
v[0] = padd<Packet2d>(vecs[0], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[0]), reinterpret_cast<Packet4ui>(vecs[0]), 8)));
v[1] = padd<Packet2d>(vecs[1], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[1]), reinterpret_cast<Packet4ui>(vecs[1]), 8)));
sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v[0]), reinterpret_cast<Packet4ui>(v[1]), 8));
return sum;
}
// Other reduction functions:
// mul
template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
@ -910,21 +871,6 @@ template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
return static_cast<float>(first);
}
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
PacketBlock<Packet4f,4> transpose;
transpose.packet[0] = vecs[0];
transpose.packet[1] = vecs[1];
transpose.packet[2] = vecs[2];
transpose.packet[3] = vecs[3];
ptranspose(transpose);
Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
sum = padd(sum, transpose.packet[2]);
sum = padd(sum, transpose.packet[3]);
return sum;
}
template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
{
// Return predux_mul<Packet2d> of the subvectors product
@ -1106,34 +1052,6 @@ template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
return pfirst(sum);
}
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
Packet4f v[4], sum[4];
// It's easier and faster to transpose then add as columns
// Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
// Do the transpose, first set of moves
v[0] = vec_mergeh(vecs[0], vecs[2]);
v[1] = vec_mergel(vecs[0], vecs[2]);
v[2] = vec_mergeh(vecs[1], vecs[3]);
v[3] = vec_mergel(vecs[1], vecs[3]);
// Get the resulting vectors
sum[0] = vec_mergeh(v[0], v[2]);
sum[1] = vec_mergel(v[0], v[2]);
sum[2] = vec_mergeh(v[1], v[3]);
sum[3] = vec_mergel(v[1], v[3]);
// Now do the summation:
// Lines 0+1
sum[0] = padd<Packet4f>(sum[0], sum[1]);
// Lines 2+3
sum[1] = padd<Packet4f>(sum[2], sum[3]);
// Add the results
sum[0] = padd<Packet4f>(sum[0], sum[1]);
return sum[0];
}
// Other reduction functions:
// mul
template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)

13
test/packetmath.cpp
View File

@ -288,19 +288,6 @@ template<typename Scalar,typename Packet> void packetmath()
ref[0] *= data1[i];
VERIFY(internal::isApprox(ref[0], internal::predux_mul(internal::pload<Packet>(data1))) && "internal::predux_mul");
if (PacketTraits::HasReduxp)
{
for (int j=0; j<PacketSize; ++j)
{
ref[j] = Scalar(0);
for (int i=0; i<PacketSize; ++i)
ref[j] += data1[i+j*PacketSize];
packets[j] = internal::pload<Packet>(data1+j*PacketSize);
}
internal::pstore(data2, internal::preduxp(packets));
VERIFY(test::areApproxAbs(ref, data2, PacketSize, refvalue) && "internal::preduxp");
}
for (int i=0; i<PacketSize; ++i)
ref[i] = data1[PacketSize-i-1];
internal::pstore(data2, internal::preverse(internal::pload<Packet>(data1)));