(So the pages are re-done …)



The bug is because NR is BlockSize<T>::NR divided by the vector length.  So this
should work

//---------------------------------------------------------------------------------
template <typename Index, typename T>
typename std::enable_if<std::is_floating_point<T>::value,
void>::type
ugemm(Index kc, T alpha,
               const T *A, const T *B,
     T beta,
     T *C, Index incRowC, Index incColC)
{
   A = (const T*) __builtin_assume_aligned (A, 128);
   B = (const T*) __builtin_assume_aligned (B, 128);
   static const unsigned div = 32/sizeof(T);
   static const Index MR = BlockSize<T>::MR;
   static const Index NR = BlockSize<T>::NR/div;

   typedef T vx __attribute__((vector_size (32)));

   vx P[MR*NR] __attribute__ ((aligned (128))) = {};
   const vx *B_ = (vx *)B;
   for (Index l=0; l<kc; ++l) {
       for (Index i=0; i<MR; ++i) {
           for (Index j=0; j<(NR); ++j) {
               P[i * NR + j] += A[i]*B_[j];
           }
       }
       A += MR;
       B_ += NR;
   }

   T *P_ = (T *)P;
   for (Index j=0; j<NR*div; ++j) {
       for (Index i=0; i<MR; ++i) {
           C[i*incRowC+j*incColC] *= beta;
           C[i*incRowC+j*incColC] += alpha*P_[i * NR * div + j];
       }
   }
}
//---------------------------------------------------------------------------------

However, as Karl pointed out it also should treat the case beta==0
as a special case.

But I don’t get any speedup.  For the performance relevant is
mainly the loop

   for (Index l=0; l<kc; ++l) {
       for (Index i=0; i<MR; ++i) {
           for (Index j=0; j<(NR); ++j) {
               P[i * NR + j] += A[i]*B_[j];
           }
       }
       A += MR;
       B_ += NR;
   }

From my experience, only if this is optimized to the extrem a little bit
extra performance can be achieved by treating some special cases for
the update

   T *P_ = (T *)P;
   for (Index j=0; j<NR*div; ++j) {
       for (Index i=0; i<MR; ++i) {
           C[i*incRowC+j*incColC] *= beta;
           C[i*incRowC+j*incColC] += alpha*P_[i * NR * div + j];
       }
   }

The main reason my code looks weird is that I also want to use it
for teaching.  The way it is coded you easily can map it forth and
back with the assembly code.


new kernel:

template <typename Index, typename T>
typename std::enable_if<std::is_floating_point<T>::value,
void>::type
ugemm(Index kc, T alpha,
               const T *A, const T *B,
     T beta,
     T *C, Index incRowC, Index incColC)
{
   A = (const T*) __builtin_assume_aligned (A, 128);
   B = (const T*) __builtin_assume_aligned (B, 128);
   static const unsigned div = 32/sizeof(T);
   static const Index MR = BlockSize<T>::MR;
   static const Index NR = BlockSize<T>::NR/div;

   typedef T vx __attribute__((vector_size (32)));

   vx P[MR*NR] __attribute__ ((aligned (128))) = {};
   const vx *B_ = (vx *)B;
   for (Index l=0; l<kc; ++l) {
       for (Index i=0; i<MR; ++i) {
           for (Index j=0; j<(NR); ++j) {
               P[i * NR + j] += A[i]*B_[j];
           }
       }
       A += MR;
       B_ += NR;
   }

   T *P_ = (T *)P;
   for (Index j=0; j<NR; ++j) {
       for (Index i=0; i<MR; ++i) {
           C[i*incRowC+j*incColC] *= beta;
           C[i*incRowC+j*incColC] += alpha*P_[i * NR + j];
       }
   }
}