From 6bc1aaaac721349e3a4f344620beb78080fef343 Mon Sep 17 00:00:00 2001
From: tensor-tang <tangjian03@baidu.com>
Date: Tue, 10 Jul 2018 11:00:44 +0800
Subject: [PATCH] refine the ColMajor replacement

---
 paddle/fluid/operators/math/blas_impl.h         |  9 +++------
 .../fluid/operators/math/math_function_test.cc  | 17 +++++++++--------
 2 files changed, 12 insertions(+), 14 deletions(-)

diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h
index 6f188636e..bb2a07667 100644
--- a/paddle/fluid/operators/math/blas_impl.h
+++ b/paddle/fluid/operators/math/blas_impl.h
@@ -168,6 +168,9 @@ void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
                                             CBLAS_TRANSPOSE transB, int M,
                                             int N, int K, T alpha, const T *A,
                                             const T *B, T beta, T *C) const {
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
+  int ldc = N;
 #ifdef PADDLE_WITH_LIBXSMM
   if (M * N * K < 128 * 128 * 128 && transA == CblasNoTrans &&
       transB == CblasNoTrans) {
@@ -175,16 +178,10 @@ void Blas<platform::CPUDeviceContext>::GEMM(CBLAS_TRANSPOSE transA,
     // Note: SMM use ColMajor
     const char transa = 'N';
     const char transb = 'N';
-    const int lda = M;
-    const int ldb = K;
-    const int ldc = M;
     CBlas<T>::SMM_GEMM(&transa, &transb, &N, &M, &K, &alpha, B, &ldb, A, &lda,
                        &beta, C, &ldc);
   } else {
 #endif
-    int lda = (transA == CblasNoTrans) ? K : M;
-    int ldb = (transB == CblasNoTrans) ? N : K;
-    int ldc = N;
     CBlas<T>::GEMM(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B,
                    ldb, beta, C, ldc);
 #ifdef PADDLE_WITH_LIBXSMM
diff --git a/paddle/fluid/operators/math/math_function_test.cc b/paddle/fluid/operators/math/math_function_test.cc
index 71103be49..078dd448c 100644
--- a/paddle/fluid/operators/math/math_function_test.cc
+++ b/paddle/fluid/operators/math/math_function_test.cc
@@ -75,26 +75,25 @@ void MklSmmCompare(int m, int n, int k) {
   for (int i = 0; i < mat_b.numel(); ++i) {
     B[i] = static_cast<T>(i);
   }
+  // lda,ldb,ldc follow RowMajor
+  int lda = k;
+  int ldb = n;
+  int ldc = n;
 
-  auto smm = [&, m, n, k, alpha, beta]() {
+  auto smm = [&, m, n, k, lda, ldb, ldc, alpha, beta]() {
     const char transa = 'N';
     const char transb = 'N';
-    const int lda = m;
-    const int ldb = k;
-    const int ldc = m;
     paddle::operators::math::CBlas<T>::SMM_GEMM(&transa, &transb, &n, &m, &k,
                                                 &alpha, B, &ldb, A, &lda, &beta,
                                                 CSMM, &ldc);
   };
 
-  auto mkl = [&, m, n, k, alpha, beta]() {
-    int lda = k;
-    int ldb = n;
-    int ldc = n;
+  auto mkl = [&, m, n, k, lda, ldb, ldc, alpha, beta]() {
     paddle::operators::math::CBlas<T>::GEMM(CblasRowMajor, CblasNoTrans,
                                             CblasNoTrans, m, n, k, alpha, A,
                                             lda, B, ldb, beta, CMKL, ldc);
   };
+
   smm();
   mkl();
   ASSERT_EQ(mat_c_mkl.numel(), mat_c_smm.numel());
@@ -105,6 +104,8 @@ void MklSmmCompare(int m, int n, int k) {
 TEST(math_function, gemm_mkl_vs_smm) {
   MklSmmCompare<float>(1, 2, 3);
   MklSmmCompare<double>(1, 2, 3);
+  MklSmmCompare<float>(3, 2, 1);
+  MklSmmCompare<double>(3, 2, 1);
   MklSmmCompare<float>(3, 8, 5);
   MklSmmCompare<double>(3, 8, 5);
 }
-- 
GitLab