Gemmer::sgemm not match caffe_cpu_gemm<float> completely

[victorygogogo]

Gemmer::sgemm not match caffe_cpu_gemm completely ,
when you test a model,you can found that the loss is iscrease.
then I found that Gemmer::sgemm not support the transformed mat.

can baidu solve it ?

[cocodark]

Is the 'loss' you mentioned the accurancy loss cased by quantification?

[victorygogogo]

I mean if the input matrix need transpose ,I use this function cause a wrong result ,may cause the loss is bigger. do you knonw the function on "caffe caffe_cpu_gemm"
void caffe_cpu_gemm(const CBLAS_TRANSPOSE TransA,
const CBLAS_TRANSPOSE TransB, const int M, const int N, const int K,
const float alpha, const float* A, const float* B, const float beta,
float* C) {
LOG(INFO) << "Running for caffe_cpu_gemm trans start" ;
int lda = (TransA == CblasNoTrans) ? K : M;
int ldb = (TransB == CblasNoTrans) ? N : K;
cblas_sgemm(CblasRowMajor, TransA, TransB, M, N, K, alpha, A, lda, B,
ldb, beta, C, N);
}

[cocodark]

Our Gemmer is not support transposing matrix indeed, we put transpose procedure in the conversion of model, in order to accelerate matrix manipulation. @victorygogogo

[cocodark]

`/**

• transpose matrix in advance

• @param data

• @param shape

• @return
  */
  float *trans_matrix(const float *data, vector shape) {

  int m = shape[0];
  int n = shape[1];

  float *trans = new float[m * n];

  for (int i = 0; i < n; ++i) {
  for (int j = 0; j < m; ++j) {

       trans[i * m + j] = data[j * n + i];
   }
  

  }
  return trans;

}` in caffe2mdl.cpp @victorygogogo

[victorygogogo]

OK ,thank you

[cocodark]

You're welcome.

[victorygogogo]

I found that ,I use this function ,it is slower than the openblas lib of gemm.by the way ,I use neon on a phone.

[cocodark]

You found our gemm is slower than openblas?

[wangshankun]

openblas considered the cache control&reuse，share job，multithreading , that three factors speed up the gemm

[victorygogogo]

@cocodark yes , I use neon ,but some matrix needs transpose,so I did transpose in this function to match the "cblas_sgemm" function. even if not use transpose ,I test 30 times cost time is about6.7s, if use the cblas_sgemm,it costs about 2.7s.

[cocodark]

@victorygogogo ,excellent research work, currently our gemm is accelerated by neon, we'll try the tricks mentioned by @wangshankun ,such as cache control&reuse 、 multithreading to make it faster.If you are interested in gemm optimization work, code contributions will be appreciated.