|| || || || || ||

---

MatNorm2DistSSE

template<class ArrayT> ArrayT* Impala::Core::Matrix::MatNorm2DistSSE ( ArrayT *  aT, ArrayT *  b   ) [inline] Definition at line 121 of file MatNorm2DistSSE.h. References ILOG_DEBUG, ILOG_ERROR, ILOG_VAR, Impala::IntAlignUp(), MatE(), MatMul(), MatNrCol(), MatNrRow(), MatPadZeros(), POINTER_ALIGNED, SIMD_MatSquareAndSumAxis0(), SIMD_MatSquareAndSumAxis1(), Impala::Timer::SplitTime(), and Impala::Timer::SplitTimeStr(). Referenced by MatNorm2DistInternal(). { ILOG_VAR(Core.Matrix.MatNorm2DistSSE); if (MatNrCol(aT) != MatNrRow(b)) { ILOG_ERROR("MatNorm2DistSSE operands: dimensionality problem"); } if(!POINTER_ALIGNED(MatE(aT, 0, 0))) { ILOG_ERROR("Pointer alignment error aT"); } if(!POINTER_ALIGNED(MatE(b, 0, 0))) { ILOG_ERROR("Pointer alignment error b"); } Timer timer; int alignCount = 4; if(sizeof(typename ArrayT::StorType) == 8) alignCount = 2; // aa = sum(multiply(a,a), axis=0) # aa = sum(a.*a,1) ArrayT* aaT = SIMD_MatSquareAndSumAxis1(aT); ILOG_DEBUG("aaT: " << timer.SplitTimeStr()); // bb = sum(multiply(b,b), axis=0) # bb = sum(b.*b,1) // this operation ensures that later on the number of columns is always // a multiple of 2 inside bb ArrayT* bb = SIMD_MatSquareAndSumAxis0(b, alignCount); ILOG_DEBUG("bb: " << timer.SplitTimeStr()); // ab = dot(transpose(a),b) # ab = a'*b //ArrayT* aT = MatTranspose(a); //ILOG_DEBUG("transpose: " << timer.SplitTimeStr()); ILOG_DEBUG("aT " << MatNrRow(aT) << " " << MatNrCol(aT)); ILOG_DEBUG("b " << MatNrRow(b) << " " << MatNrCol(b)); ArrayT* paddedB = MatPadZeros(b, MatNrRow(b), IntAlignUp(MatNrCol(b), alignCount)); ArrayT* ab = MatMul(aT, paddedB); delete paddedB; ILOG_DEBUG("matmul: " << timer.SplitTimeStr()); // return sqrt(abs(transpose(repmat(aa,bb.shape[0],1)) + repmat(bb,aa.shape[0],1) - 2*ab)) // #d = sqrt(abs(repmat(aa',[1 size(bb,2)]) + repmat(bb,[size(aa,2) 1]) - 2*ab)); //ILOG_DEBUG("aaT " << MatNrRow(aaT) << " " << MatNrCol(aaT)); //ILOG_DEBUG("bb " << MatNrRow(bb) << " " << MatNrCol(bb)); if(sizeof(typename ArrayT::StorType) == 8) { #pragma omp parallel for for (int i=0 ; i<MatNrRow(aaT) ; i++) { const __m128d tmp = _mm_set1_pd(*MatE(aaT, i, 0)); const __m128d minusTwo = _mm_set1_pd(-2.0); const UInt64 mask = 0x7FFFFFFFFFFFFFFFL; const __m128d absMask = _mm_set1_pd(*(double*)(&mask)); __m128d* baseBB = (__m128d*)(MatE(bb, 0, 0)); __m128d* baseAB = (__m128d*)(MatE(ab, i, 0)); const int SSELength = MatNrCol(bb) / 2; for(int j = 0; j < SSELength; j++) { __m128d intermediate = _mm_add_pd(_mm_add_pd(baseBB[j], tmp), _mm_mul_pd(minusTwo, baseAB[j])); baseAB[j] = _mm_sqrt_pd(_mm_and_pd(intermediate, absMask)); } } } else { #pragma omp parallel for for (int i=0 ; i<MatNrRow(aaT) ; i++) { const __m128 tmp = _mm_set1_ps(*MatE(aaT, i, 0)); const __m128 minusTwo = _mm_set1_ps(-2.0); const UInt32 mask = 0x7FFFFFFF; const __m128 absMask = _mm_set1_ps(*(Real32*)(&mask)); __m128* baseBB = (__m128*)(MatE(bb, 0, 0)); __m128* baseAB = (__m128*)(MatE(ab, i, 0)); const int SSELength = MatNrCol(bb) / 4; for(int j = 0; j < SSELength; j++) { __m128 intermediate = _mm_add_ps(_mm_add_ps(baseBB[j], tmp), _mm_mul_ps(minusTwo, baseAB[j])); baseAB[j] = _mm_sqrt_ps(_mm_and_ps(intermediate, absMask)); } } } ILOG_DEBUG("sqrt(abs(aa + bb - 2ab)): " << timer.SplitTimeStr()); delete bb; delete aaT; ILOG_DEBUG("delete: " << timer.SplitTimeStr()); //WriteRaw(a, "matrix_a.raw", &Util::Database::GetInstance(), 1); //WriteRaw(b, "matrix_b.raw", &Util::Database::GetInstance(), 1); //WriteRaw(repAAT, "matrix_c.raw", &Util::Database::GetInstance(), 1); //ILOG_INFO("cpu-matnorm2dist-total: " << timer.SplitTime()); ILOG_DEBUG(timer.SplitTime() << " (cpu-matnorm2dist-total)"); return ab; } Here is the call graph for this function:

---