portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

Improvements to avx for invertible to avoid latent bug

Browse Source
This commit is contained in:
paboyle 2017-02-16 23:52:44 +00:00
parent 8a29c16bde
commit f246fe3304
2 changed files with 21 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
lib/simd/Grid_avx.h
View File

@ -474,16 +474,31 @@ namespace Optimization {
struct Exchange{ struct Exchange{
// 3210 ordering // 3210 ordering
static inline void Exchange0(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){ static inline void Exchange0(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){
//Invertible
//AB CD -> AC BD
//AC BD -> AB CD
out1= _mm256_permute2f128_ps(in1,in2,0x20); out1= _mm256_permute2f128_ps(in1,in2,0x20);
out2= _mm256_permute2f128_ps(in1,in2,0x31); out2= _mm256_permute2f128_ps(in1,in2,0x31);
}; };
static inline void Exchange1(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){ static inline void Exchange1(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){
//Invertible
// ABCD EFGH ->ABEF CDGH
// ABEF CDGH ->ABCD EFGH
out1= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(1,0,1,0)); out1= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(1,0,1,0));
out2= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,2,3,2)); out2= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,2,3,2));
}; };
static inline void Exchange2(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){ static inline void Exchange2(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){
out1= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0)); // Invertible ?
out2= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1)); // ABCD EFGH -> ACEG BDFH
// ACEG BDFH -> AEBF CGDH
// out1= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0));
// out2= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1));
// Bollocks; need
// AECG BFDH -> ABCD EFGH
out1= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(2,0,2,0)); /*ACEG*/
out2= _mm256_shuffle_ps(in1,in2,_MM_SELECT_FOUR_FOUR(3,1,3,1)); /*BDFH*/
out1= _mm256_shuffle_ps(out1,out1,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
out2= _mm256_shuffle_ps(out2,out2,_MM_SELECT_FOUR_FOUR(3,1,2,0)); /*AECG*/
}; };
static inline void Exchange3(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){ static inline void Exchange3(__m256 &out1,__m256 &out2,__m256 in1,__m256 in2){
assert(0); assert(0);

6
tests/Test_simd.cc
View File

@ -419,8 +419,10 @@ void ExchangeTester(const functor &func)
assert(found==1); assert(found==1);
} }
// for(int i=0;i<Nsimd;i++){ for(int i=0;i<Nsimd;i++){
// std::cout << " i "<< i<<" test1"<<test1[i]<<" "<<input1[i]<<std::endl; assert(test1[i]==input1[i]);
assert(test2[i]==input2[i]);
}// std::cout << " i "<< i<<" test1"<<test1[i]<<" "<<input1[i]<<std::endl;
// std::cout << " i "<< i<<" test2"<<test2[i]<<" "<<input2[i]<<std::endl; // std::cout << " i "<< i<<" test2"<<test2[i]<<" "<<input2[i]<<std::endl;
// } // }
} }