1 | // This file is part of Eigen, a lightweight C++ template library |
---|
2 | // for linear algebra. |
---|
3 | // |
---|
4 | // Copyright (C) 2008 Konstantinos Margaritis <markos@codex.gr> |
---|
5 | // |
---|
6 | // This Source Code Form is subject to the terms of the Mozilla |
---|
7 | // Public License v. 2.0. If a copy of the MPL was not distributed |
---|
8 | // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. |
---|
9 | |
---|
10 | #ifndef EIGEN_PACKET_MATH_ALTIVEC_H |
---|
11 | #define EIGEN_PACKET_MATH_ALTIVEC_H |
---|
12 | |
---|
13 | namespace Eigen { |
---|
14 | |
---|
15 | namespace internal { |
---|
16 | |
---|
17 | #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD |
---|
18 | #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4 |
---|
19 | #endif |
---|
20 | |
---|
21 | #ifndef EIGEN_HAS_FUSE_CJMADD |
---|
22 | #define EIGEN_HAS_FUSE_CJMADD 1 |
---|
23 | #endif |
---|
24 | |
---|
25 | // NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16 |
---|
26 | #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS |
---|
27 | #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 |
---|
28 | #endif |
---|
29 | |
---|
30 | typedef __vector float Packet4f; |
---|
31 | typedef __vector int Packet4i; |
---|
32 | typedef __vector unsigned int Packet4ui; |
---|
33 | typedef __vector __bool int Packet4bi; |
---|
34 | typedef __vector short int Packet8i; |
---|
35 | typedef __vector unsigned char Packet16uc; |
---|
36 | |
---|
37 | // We don't want to write the same code all the time, but we need to reuse the constants |
---|
38 | // and it doesn't really work to declare them global, so we define macros instead |
---|
39 | |
---|
40 | #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \ |
---|
41 | Packet4f p4f_##NAME = (Packet4f) vec_splat_s32(X) |
---|
42 | |
---|
43 | #define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \ |
---|
44 | Packet4i p4i_##NAME = vec_splat_s32(X) |
---|
45 | |
---|
46 | #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ |
---|
47 | Packet4f p4f_##NAME = pset1<Packet4f>(X) |
---|
48 | |
---|
49 | #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ |
---|
50 | Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int>(X)) |
---|
51 | |
---|
52 | #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ |
---|
53 | Packet4i p4i_##NAME = pset1<Packet4i>(X) |
---|
54 | |
---|
55 | #define DST_CHAN 1 |
---|
56 | #define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride)) |
---|
57 | |
---|
58 | // Define global static constants: |
---|
59 | static Packet4f p4f_COUNTDOWN = { 3.0, 2.0, 1.0, 0.0 }; |
---|
60 | static Packet4i p4i_COUNTDOWN = { 3, 2, 1, 0 }; |
---|
61 | static Packet16uc p16uc_REVERSE = {12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3}; |
---|
62 | static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0); |
---|
63 | static Packet16uc p16uc_DUPLICATE = {0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7}; |
---|
64 | |
---|
65 | static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); |
---|
66 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); |
---|
67 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); |
---|
68 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); |
---|
69 | static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); |
---|
70 | static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); |
---|
71 | static Packet4f p4f_ZERO_ = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); |
---|
72 | |
---|
73 | template<> struct packet_traits<float> : default_packet_traits |
---|
74 | { |
---|
75 | typedef Packet4f type; |
---|
76 | enum { |
---|
77 | Vectorizable = 1, |
---|
78 | AlignedOnScalar = 1, |
---|
79 | size=4, |
---|
80 | |
---|
81 | // FIXME check the Has* |
---|
82 | HasSin = 0, |
---|
83 | HasCos = 0, |
---|
84 | HasLog = 0, |
---|
85 | HasExp = 0, |
---|
86 | HasSqrt = 0 |
---|
87 | }; |
---|
88 | }; |
---|
89 | template<> struct packet_traits<int> : default_packet_traits |
---|
90 | { |
---|
91 | typedef Packet4i type; |
---|
92 | enum { |
---|
93 | // FIXME check the Has* |
---|
94 | Vectorizable = 1, |
---|
95 | AlignedOnScalar = 1, |
---|
96 | size=4 |
---|
97 | }; |
---|
98 | }; |
---|
99 | |
---|
100 | template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; }; |
---|
101 | template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; }; |
---|
102 | /* |
---|
103 | inline std::ostream & operator <<(std::ostream & s, const Packet4f & v) |
---|
104 | { |
---|
105 | union { |
---|
106 | Packet4f v; |
---|
107 | float n[4]; |
---|
108 | } vt; |
---|
109 | vt.v = v; |
---|
110 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; |
---|
111 | return s; |
---|
112 | } |
---|
113 | |
---|
114 | inline std::ostream & operator <<(std::ostream & s, const Packet4i & v) |
---|
115 | { |
---|
116 | union { |
---|
117 | Packet4i v; |
---|
118 | int n[4]; |
---|
119 | } vt; |
---|
120 | vt.v = v; |
---|
121 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; |
---|
122 | return s; |
---|
123 | } |
---|
124 | |
---|
125 | inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v) |
---|
126 | { |
---|
127 | union { |
---|
128 | Packet4ui v; |
---|
129 | unsigned int n[4]; |
---|
130 | } vt; |
---|
131 | vt.v = v; |
---|
132 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; |
---|
133 | return s; |
---|
134 | } |
---|
135 | |
---|
136 | inline std::ostream & operator <<(std::ostream & s, const Packetbi & v) |
---|
137 | { |
---|
138 | union { |
---|
139 | Packet4bi v; |
---|
140 | unsigned int n[4]; |
---|
141 | } vt; |
---|
142 | vt.v = v; |
---|
143 | s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3]; |
---|
144 | return s; |
---|
145 | } |
---|
146 | */ |
---|
147 | template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { |
---|
148 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html |
---|
149 | float EIGEN_ALIGN16 af[4]; |
---|
150 | af[0] = from; |
---|
151 | Packet4f vc = vec_ld(0, af); |
---|
152 | vc = vec_splat(vc, 0); |
---|
153 | return vc; |
---|
154 | } |
---|
155 | |
---|
156 | template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { |
---|
157 | int EIGEN_ALIGN16 ai[4]; |
---|
158 | ai[0] = from; |
---|
159 | Packet4i vc = vec_ld(0, ai); |
---|
160 | vc = vec_splat(vc, 0); |
---|
161 | return vc; |
---|
162 | } |
---|
163 | |
---|
164 | template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return vec_add(pset1<Packet4f>(a), p4f_COUNTDOWN); } |
---|
165 | template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return vec_add(pset1<Packet4i>(a), p4i_COUNTDOWN); } |
---|
166 | |
---|
167 | template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_add(a,b); } |
---|
168 | template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_add(a,b); } |
---|
169 | |
---|
170 | template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_sub(a,b); } |
---|
171 | template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_sub(a,b); } |
---|
172 | |
---|
173 | template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return psub<Packet4f>(p4f_ZERO, a); } |
---|
174 | template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return psub<Packet4i>(p4i_ZERO, a); } |
---|
175 | |
---|
176 | template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b,p4f_ZERO); } |
---|
177 | /* Commented out: it's actually slower than processing it scalar |
---|
178 | * |
---|
179 | template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) |
---|
180 | { |
---|
181 | // Detailed in: http://freevec.org/content/32bit_signed_integer_multiplication_altivec |
---|
182 | //Set up constants, variables |
---|
183 | Packet4i a1, b1, bswap, low_prod, high_prod, prod, prod_, v1sel; |
---|
184 | |
---|
185 | // Get the absolute values |
---|
186 | a1 = vec_abs(a); |
---|
187 | b1 = vec_abs(b); |
---|
188 | |
---|
189 | // Get the signs using xor |
---|
190 | Packet4bi sgn = (Packet4bi) vec_cmplt(vec_xor(a, b), p4i_ZERO); |
---|
191 | |
---|
192 | // Do the multiplication for the asbolute values. |
---|
193 | bswap = (Packet4i) vec_rl((Packet4ui) b1, (Packet4ui) p4i_MINUS16 ); |
---|
194 | low_prod = vec_mulo((Packet8i) a1, (Packet8i)b1); |
---|
195 | high_prod = vec_msum((Packet8i) a1, (Packet8i) bswap, p4i_ZERO); |
---|
196 | high_prod = (Packet4i) vec_sl((Packet4ui) high_prod, (Packet4ui) p4i_MINUS16); |
---|
197 | prod = vec_add( low_prod, high_prod ); |
---|
198 | |
---|
199 | // NOR the product and select only the negative elements according to the sign mask |
---|
200 | prod_ = vec_nor(prod, prod); |
---|
201 | prod_ = vec_sel(p4i_ZERO, prod_, sgn); |
---|
202 | |
---|
203 | // Add 1 to the result to get the negative numbers |
---|
204 | v1sel = vec_sel(p4i_ZERO, p4i_ONE, sgn); |
---|
205 | prod_ = vec_add(prod_, v1sel); |
---|
206 | |
---|
207 | // Merge the results back to the final vector. |
---|
208 | prod = vec_sel(prod, prod_, sgn); |
---|
209 | |
---|
210 | return prod; |
---|
211 | } |
---|
212 | */ |
---|
213 | template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) |
---|
214 | { |
---|
215 | Packet4f t, y_0, y_1, res; |
---|
216 | |
---|
217 | // Altivec does not offer a divide instruction, we have to do a reciprocal approximation |
---|
218 | y_0 = vec_re(b); |
---|
219 | |
---|
220 | // Do one Newton-Raphson iteration to get the needed accuracy |
---|
221 | t = vec_nmsub(y_0, b, p4f_ONE); |
---|
222 | y_1 = vec_madd(y_0, t, y_0); |
---|
223 | |
---|
224 | res = vec_madd(a, y_1, p4f_ZERO); |
---|
225 | return res; |
---|
226 | } |
---|
227 | |
---|
228 | template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) |
---|
229 | { eigen_assert(false && "packet integer division are not supported by AltiVec"); |
---|
230 | return pset1<Packet4i>(0); |
---|
231 | } |
---|
232 | |
---|
233 | // for some weird raisons, it has to be overloaded for packet of integers |
---|
234 | template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a, b, c); } |
---|
235 | template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } |
---|
236 | |
---|
237 | template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); } |
---|
238 | template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); } |
---|
239 | |
---|
240 | template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); } |
---|
241 | template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); } |
---|
242 | |
---|
243 | // Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics |
---|
244 | template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); } |
---|
245 | template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); } |
---|
246 | |
---|
247 | template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); } |
---|
248 | template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); } |
---|
249 | |
---|
250 | template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); } |
---|
251 | template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); } |
---|
252 | |
---|
253 | template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); } |
---|
254 | template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); } |
---|
255 | |
---|
256 | template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } |
---|
257 | template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return vec_ld(0, from); } |
---|
258 | |
---|
259 | template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) |
---|
260 | { |
---|
261 | EIGEN_DEBUG_ALIGNED_LOAD |
---|
262 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html |
---|
263 | Packet16uc MSQ, LSQ; |
---|
264 | Packet16uc mask; |
---|
265 | MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword |
---|
266 | LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword |
---|
267 | mask = vec_lvsl(0, from); // create the permute mask |
---|
268 | return (Packet4f) vec_perm(MSQ, LSQ, mask); // align the data |
---|
269 | |
---|
270 | } |
---|
271 | template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) |
---|
272 | { |
---|
273 | EIGEN_DEBUG_ALIGNED_LOAD |
---|
274 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html |
---|
275 | Packet16uc MSQ, LSQ; |
---|
276 | Packet16uc mask; |
---|
277 | MSQ = vec_ld(0, (unsigned char *)from); // most significant quadword |
---|
278 | LSQ = vec_ld(15, (unsigned char *)from); // least significant quadword |
---|
279 | mask = vec_lvsl(0, from); // create the permute mask |
---|
280 | return (Packet4i) vec_perm(MSQ, LSQ, mask); // align the data |
---|
281 | } |
---|
282 | |
---|
283 | template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) |
---|
284 | { |
---|
285 | Packet4f p; |
---|
286 | if((ptrdiff_t(&from) % 16) == 0) p = pload<Packet4f>(from); |
---|
287 | else p = ploadu<Packet4f>(from); |
---|
288 | return vec_perm(p, p, p16uc_DUPLICATE); |
---|
289 | } |
---|
290 | template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) |
---|
291 | { |
---|
292 | Packet4i p; |
---|
293 | if((ptrdiff_t(&from) % 16) == 0) p = pload<Packet4i>(from); |
---|
294 | else p = ploadu<Packet4i>(from); |
---|
295 | return vec_perm(p, p, p16uc_DUPLICATE); |
---|
296 | } |
---|
297 | |
---|
298 | template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } |
---|
299 | template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vec_st(from, 0, to); } |
---|
300 | |
---|
301 | template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) |
---|
302 | { |
---|
303 | EIGEN_DEBUG_UNALIGNED_STORE |
---|
304 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html |
---|
305 | // Warning: not thread safe! |
---|
306 | Packet16uc MSQ, LSQ, edges; |
---|
307 | Packet16uc edgeAlign, align; |
---|
308 | |
---|
309 | MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword |
---|
310 | LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword |
---|
311 | edgeAlign = vec_lvsl(0, to); // permute map to extract edges |
---|
312 | edges=vec_perm(LSQ,MSQ,edgeAlign); // extract the edges |
---|
313 | align = vec_lvsr( 0, to ); // permute map to misalign data |
---|
314 | MSQ = vec_perm(edges,(Packet16uc)from,align); // misalign the data (MSQ) |
---|
315 | LSQ = vec_perm((Packet16uc)from,edges,align); // misalign the data (LSQ) |
---|
316 | vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first |
---|
317 | vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part |
---|
318 | } |
---|
319 | template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) |
---|
320 | { |
---|
321 | EIGEN_DEBUG_UNALIGNED_STORE |
---|
322 | // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html |
---|
323 | // Warning: not thread safe! |
---|
324 | Packet16uc MSQ, LSQ, edges; |
---|
325 | Packet16uc edgeAlign, align; |
---|
326 | |
---|
327 | MSQ = vec_ld(0, (unsigned char *)to); // most significant quadword |
---|
328 | LSQ = vec_ld(15, (unsigned char *)to); // least significant quadword |
---|
329 | edgeAlign = vec_lvsl(0, to); // permute map to extract edges |
---|
330 | edges=vec_perm(LSQ, MSQ, edgeAlign); // extract the edges |
---|
331 | align = vec_lvsr( 0, to ); // permute map to misalign data |
---|
332 | MSQ = vec_perm(edges, (Packet16uc) from, align); // misalign the data (MSQ) |
---|
333 | LSQ = vec_perm((Packet16uc) from, edges, align); // misalign the data (LSQ) |
---|
334 | vec_st( LSQ, 15, (unsigned char *)to ); // Store the LSQ part first |
---|
335 | vec_st( MSQ, 0, (unsigned char *)to ); // Store the MSQ part |
---|
336 | } |
---|
337 | |
---|
338 | template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } |
---|
339 | template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { vec_dstt(addr, DST_CTRL(2,2,32), DST_CHAN); } |
---|
340 | |
---|
341 | template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } |
---|
342 | template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int EIGEN_ALIGN16 x[4]; vec_st(a, 0, x); return x[0]; } |
---|
343 | |
---|
344 | template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) { return (Packet4f)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } |
---|
345 | template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) { return (Packet4i)vec_perm((Packet16uc)a,(Packet16uc)a, p16uc_REVERSE); } |
---|
346 | |
---|
347 | template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); } |
---|
348 | template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); } |
---|
349 | |
---|
350 | template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) |
---|
351 | { |
---|
352 | Packet4f b, sum; |
---|
353 | b = (Packet4f) vec_sld(a, a, 8); |
---|
354 | sum = vec_add(a, b); |
---|
355 | b = (Packet4f) vec_sld(sum, sum, 4); |
---|
356 | sum = vec_add(sum, b); |
---|
357 | return pfirst(sum); |
---|
358 | } |
---|
359 | |
---|
360 | template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) |
---|
361 | { |
---|
362 | Packet4f v[4], sum[4]; |
---|
363 | |
---|
364 | // It's easier and faster to transpose then add as columns |
---|
365 | // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation |
---|
366 | // Do the transpose, first set of moves |
---|
367 | v[0] = vec_mergeh(vecs[0], vecs[2]); |
---|
368 | v[1] = vec_mergel(vecs[0], vecs[2]); |
---|
369 | v[2] = vec_mergeh(vecs[1], vecs[3]); |
---|
370 | v[3] = vec_mergel(vecs[1], vecs[3]); |
---|
371 | // Get the resulting vectors |
---|
372 | sum[0] = vec_mergeh(v[0], v[2]); |
---|
373 | sum[1] = vec_mergel(v[0], v[2]); |
---|
374 | sum[2] = vec_mergeh(v[1], v[3]); |
---|
375 | sum[3] = vec_mergel(v[1], v[3]); |
---|
376 | |
---|
377 | // Now do the summation: |
---|
378 | // Lines 0+1 |
---|
379 | sum[0] = vec_add(sum[0], sum[1]); |
---|
380 | // Lines 2+3 |
---|
381 | sum[1] = vec_add(sum[2], sum[3]); |
---|
382 | // Add the results |
---|
383 | sum[0] = vec_add(sum[0], sum[1]); |
---|
384 | |
---|
385 | return sum[0]; |
---|
386 | } |
---|
387 | |
---|
388 | template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) |
---|
389 | { |
---|
390 | Packet4i sum; |
---|
391 | sum = vec_sums(a, p4i_ZERO); |
---|
392 | sum = vec_sld(sum, p4i_ZERO, 12); |
---|
393 | return pfirst(sum); |
---|
394 | } |
---|
395 | |
---|
396 | template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) |
---|
397 | { |
---|
398 | Packet4i v[4], sum[4]; |
---|
399 | |
---|
400 | // It's easier and faster to transpose then add as columns |
---|
401 | // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation |
---|
402 | // Do the transpose, first set of moves |
---|
403 | v[0] = vec_mergeh(vecs[0], vecs[2]); |
---|
404 | v[1] = vec_mergel(vecs[0], vecs[2]); |
---|
405 | v[2] = vec_mergeh(vecs[1], vecs[3]); |
---|
406 | v[3] = vec_mergel(vecs[1], vecs[3]); |
---|
407 | // Get the resulting vectors |
---|
408 | sum[0] = vec_mergeh(v[0], v[2]); |
---|
409 | sum[1] = vec_mergel(v[0], v[2]); |
---|
410 | sum[2] = vec_mergeh(v[1], v[3]); |
---|
411 | sum[3] = vec_mergel(v[1], v[3]); |
---|
412 | |
---|
413 | // Now do the summation: |
---|
414 | // Lines 0+1 |
---|
415 | sum[0] = vec_add(sum[0], sum[1]); |
---|
416 | // Lines 2+3 |
---|
417 | sum[1] = vec_add(sum[2], sum[3]); |
---|
418 | // Add the results |
---|
419 | sum[0] = vec_add(sum[0], sum[1]); |
---|
420 | |
---|
421 | return sum[0]; |
---|
422 | } |
---|
423 | |
---|
424 | // Other reduction functions: |
---|
425 | // mul |
---|
426 | template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) |
---|
427 | { |
---|
428 | Packet4f prod; |
---|
429 | prod = pmul(a, (Packet4f)vec_sld(a, a, 8)); |
---|
430 | return pfirst(pmul(prod, (Packet4f)vec_sld(prod, prod, 4))); |
---|
431 | } |
---|
432 | |
---|
433 | template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) |
---|
434 | { |
---|
435 | EIGEN_ALIGN16 int aux[4]; |
---|
436 | pstore(aux, a); |
---|
437 | return aux[0] * aux[1] * aux[2] * aux[3]; |
---|
438 | } |
---|
439 | |
---|
440 | // min |
---|
441 | template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) |
---|
442 | { |
---|
443 | Packet4f b, res; |
---|
444 | b = vec_min(a, vec_sld(a, a, 8)); |
---|
445 | res = vec_min(b, vec_sld(b, b, 4)); |
---|
446 | return pfirst(res); |
---|
447 | } |
---|
448 | |
---|
449 | template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) |
---|
450 | { |
---|
451 | Packet4i b, res; |
---|
452 | b = vec_min(a, vec_sld(a, a, 8)); |
---|
453 | res = vec_min(b, vec_sld(b, b, 4)); |
---|
454 | return pfirst(res); |
---|
455 | } |
---|
456 | |
---|
457 | // max |
---|
458 | template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) |
---|
459 | { |
---|
460 | Packet4f b, res; |
---|
461 | b = vec_max(a, vec_sld(a, a, 8)); |
---|
462 | res = vec_max(b, vec_sld(b, b, 4)); |
---|
463 | return pfirst(res); |
---|
464 | } |
---|
465 | |
---|
466 | template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) |
---|
467 | { |
---|
468 | Packet4i b, res; |
---|
469 | b = vec_max(a, vec_sld(a, a, 8)); |
---|
470 | res = vec_max(b, vec_sld(b, b, 4)); |
---|
471 | return pfirst(res); |
---|
472 | } |
---|
473 | |
---|
474 | template<int Offset> |
---|
475 | struct palign_impl<Offset,Packet4f> |
---|
476 | { |
---|
477 | static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) |
---|
478 | { |
---|
479 | if (Offset!=0) |
---|
480 | first = vec_sld(first, second, Offset*4); |
---|
481 | } |
---|
482 | }; |
---|
483 | |
---|
484 | template<int Offset> |
---|
485 | struct palign_impl<Offset,Packet4i> |
---|
486 | { |
---|
487 | static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) |
---|
488 | { |
---|
489 | if (Offset!=0) |
---|
490 | first = vec_sld(first, second, Offset*4); |
---|
491 | } |
---|
492 | }; |
---|
493 | |
---|
494 | } // end namespace internal |
---|
495 | |
---|
496 | } // end namespace Eigen |
---|
497 | |
---|
498 | #endif // EIGEN_PACKET_MATH_ALTIVEC_H |
---|