From 9e6a569cf4cbf5d31cd25fcc0a5723a0d6a8acb3 Mon Sep 17 00:00:00 2001
From: Aayush <aankit@purdue.edu>
Date: Sat, 26 Oct 2019 01:02:40 -0400
Subject: [PATCH] Patched with diff from cutlass-gpgpu-sim-2

---
 gemm-test/dgemm_tests.h |  11 +-
 gemm-test/sgemm_tests.h |   8 +-
 gemm-test/wmma_tests.h  | 333 +++++++++++++++-------------------------
 3 files changed, 130 insertions(+), 222 deletions(-)

diff --git a/gemm-test/dgemm_tests.h b/gemm-test/dgemm_tests.h
index c9e84d4..439f212 100644
--- a/gemm-test/dgemm_tests.h
+++ b/gemm-test/dgemm_tests.h
@@ -1,20 +1,19 @@
-  #ifdef DGEMM_1
   typedef cutlass::gemm::DgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor,
                                      cutlass::Shape<8, 32, 64> > DGemmTraits1;
   run_gemm<DGemmTraits1>(64, 32, 8);
-  #endif
 
-  #ifdef DGEMM_2
+
+
   typedef cutlass::gemm::DgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor,
                                      cutlass::Shape<8, 32, 64> > DGemmTraits2;
   run_gemm<DGemmTraits2>(256, 128, 64);
-  #endif
 
-  #ifdef DGEMM_3
+
+
   typedef cutlass::gemm::DgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor,
                                      cutlass::Shape<8, 64, 64> > DGemmTraits3;
   run_gemm<DGemmTraits3>(64, 64, 8);
-  #endif
+
diff --git a/gemm-test/sgemm_tests.h b/gemm-test/sgemm_tests.h
index 41693ab..239327a 100644
--- a/gemm-test/sgemm_tests.h
+++ b/gemm-test/sgemm_tests.h
@@ -1,20 +1,16 @@
-  #ifdef SGEMM_1
   typedef cutlass::gemm::SgemmTraits<cutlass::MatrixLayout::kColumnMajor,
     cutlass::MatrixLayout::kRowMajor, cutlass::Shape<8, 128, 128> >
     SgemmTraits1;
   run_gemm<SgemmTraits1>(1024, 512, 8);
-  #endif
 
-  #ifdef SGEMM_2
+
   typedef cutlass::gemm::SgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor, cutlass::Shape<8, 128, 128> >
       SgemmTraits2;
   run_gemm<SgemmTraits2>(128, 81, 1);
-  #endif
 
-  #ifdef SGEMM_3
+
   typedef cutlass::gemm::SgemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                      cutlass::MatrixLayout::kRowMajor, cutlass::Shape<8, 128, 128> >
       SgemmTraits3;
   run_gemm<SgemmTraits3>(128, 112, 8);
-  #endif
diff --git a/gemm-test/wmma_tests.h b/gemm-test/wmma_tests.h
index c3a1fdd..86e16e4 100644
--- a/gemm-test/wmma_tests.h
+++ b/gemm-test/wmma_tests.h
@@ -1,26 +1,21 @@
-  #ifdef WMMA_1
+
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits1;
   run_gemm<WmmaGemmTraits1>(16, 16, 16);
-  #endif
-  #ifdef WMMA_2
+
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits2;
   run_gemm<WmmaGemmTraits2>(16, 16, 32);
-  #endif
-  #ifdef WMMA_3
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits3;
   run_gemm<WmmaGemmTraits3>(16, 16, 16);
-  #endif
-  #ifdef WMMA_4
 
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
@@ -28,316 +23,234 @@
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits4;
   run_gemm<WmmaGemmTraits4>(16, 16, 32);
-  #endif
-  #ifdef WMMA_5
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits5;
   run_gemm<WmmaGemmTraits5>(16, 16, 16);
-  #endif
-  #ifdef WMMA_6
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits6;
   run_gemm<WmmaGemmTraits6>(16, 16, 32);
-  #endif
-  #ifdef WMMA_7
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits7;
   run_gemm<WmmaGemmTraits7>(16, 16, 16);
-  #endif
-  #ifdef WMMA_8
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits8;
   run_gemm<WmmaGemmTraits8>(16, 16, 32);
-  #endif
-  #ifdef WMMA_9
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits9;
   run_gemm<WmmaGemmTraits9>(32, 32, 32);
-  #endif
-  #ifdef WMMA_10
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits10;
   run_gemm<WmmaGemmTraits10>(32, 32, 32);
-  #endif
-  #ifdef WMMA_11
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits11;
   run_gemm<WmmaGemmTraits11>(32, 32, 32);
-  #endif
-  #ifdef WMMA_12
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits12;
   run_gemm<WmmaGemmTraits12>(32, 32, 32);
-  #endif
-  #ifdef WMMA_13
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<32, 128, 128> >
+                                        cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits13;
   run_gemm<WmmaGemmTraits13>(128, 128, 128);
-  #endif
-  #ifdef WMMA_14
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<32, 128, 128> >
+                                        cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits14;
   run_gemm<WmmaGemmTraits14>(128, 128, 128);
-  #endif
-  #ifdef WMMA_15
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<32, 128, 128> >
+                                        cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits15;
   run_gemm<WmmaGemmTraits15>(128, 128, 128);
-  #endif
-  #ifdef WMMA_16
-
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<32, 128, 128> >
+                                        cutlass::Shape<32, 16, 16> >
       WmmaGemmTraits16;
   run_gemm<WmmaGemmTraits16>(128, 128, 128);
-  #endif
-  #ifdef WMMA_17
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 128, 128> >
       WmmaGemmTraits17;
   run_gemm<WmmaGemmTraits17>(256, 256, 128);
-  #endif
-  #ifdef WMMA_18
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 128, 128> >
       WmmaGemmTraits18;
   run_gemm<WmmaGemmTraits18>(256, 256, 128);
-  #endif
-  #ifdef WMMA_19
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<32, 128, 128> >
       WmmaGemmTraits19;
   run_gemm<WmmaGemmTraits19>(256, 256, 128);
-  #endif
-  #ifdef WMMA_20
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<32, 128, 128> >
       WmmaGemmTraits20;
   run_gemm<WmmaGemmTraits20>(256, 256, 128);
-  #endif
-  #ifdef WMMA_21
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<64, 128, 128> >
       WmmaGemmTraits21;
   run_gemm<WmmaGemmTraits21>(256, 256, 256);
-  #endif
-  #ifdef WMMA_22
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<64, 128, 128> >
       WmmaGemmTraits22;
   run_gemm<WmmaGemmTraits22>(256, 256, 256);
-  #endif
-  #ifdef WMMA_23
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kRowMajor,
                                         cutlass::Shape<64, 128, 128> >
       WmmaGemmTraits23;
   run_gemm<WmmaGemmTraits23>(256, 256, 256);
-  #endif
-  #ifdef WMMA_24
 
   typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
                                         cutlass::MatrixLayout::kColumnMajor,
                                         cutlass::Shape<64, 128, 128> >
       WmmaGemmTraits24;
   run_gemm<WmmaGemmTraits24>(256, 256, 256);
-  #endif
-  #ifdef WMMA_25
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits25;
-  run_gemm<WmmaGemmTraits25>(512, 512, 512);
-  #endif
-  #ifdef WMMA_26
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits26;
-  run_gemm<WmmaGemmTraits26>(512, 512, 512);
-  #endif
-  #ifdef WMMA_27
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits27;
-  run_gemm<WmmaGemmTraits27>(512, 512, 512);
-  #endif
-  #ifdef WMMA_28
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits28;
-  run_gemm<WmmaGemmTraits28>(512, 512, 512);
-  #endif
-  #ifdef WMMA_29
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits29;
-  run_gemm<WmmaGemmTraits29>(768, 768, 768);
-  #endif
-  #ifdef WMMA_30
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits30;
-  run_gemm<WmmaGemmTraits30>(768, 768, 768);
-  #endif
-  #ifdef WMMA_31
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits31;
-  run_gemm<WmmaGemmTraits31>(768, 768, 768);
-  #endif
-  #ifdef WMMA_32
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits32;
-  run_gemm<WmmaGemmTraits32>(768, 768, 768);
-  #endif
-  #ifdef WMMA_33
-
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits33;
-  run_gemm<WmmaGemmTraits33>(1024, 1024, 1024);
-  #endif
-  #ifdef WMMA_34
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits34;
-  run_gemm<WmmaGemmTraits34>(1024, 1024, 1024);
-  #endif
-  #ifdef WMMA_35
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits35;
-  run_gemm<WmmaGemmTraits35>(1024, 1024, 1024);
-  #endif
-  #ifdef WMMA_36
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits36;
-  run_gemm<WmmaGemmTraits36>(1024, 1024, 1024);
-  #endif
-  #ifdef WMMA_37
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits37;
-  run_gemm<WmmaGemmTraits37>(2048, 2048, 2048);
-  #endif
-  #ifdef WMMA_38
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits38;
-  run_gemm<WmmaGemmTraits38>(2048, 2048, 2048);
-  #endif
-  #ifdef WMMA_39
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits39;
-  run_gemm<WmmaGemmTraits39>(2048, 2048, 2048);
-  #endif
-  #ifdef WMMA_40
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits40;
-  run_gemm<WmmaGemmTraits40>(2048, 2048, 2048);
-  #endif
-  #ifdef WMMA_41
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits41;
-  run_gemm<WmmaGemmTraits41>(4096, 4096, 4096);
-  #endif
-  #ifdef WMMA_42
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits42;
-  run_gemm<WmmaGemmTraits42>(4096, 4096, 4096);
-  #endif
-  #ifdef WMMA_43
-
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits43;
-  run_gemm<WmmaGemmTraits43>(4096, 4096, 4096);
-  #endif
-  #ifdef WMMA_44
-  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
-                                        cutlass::MatrixLayout::kColumnMajor,
-                                        cutlass::Shape<64, 128, 128> >
-      WmmaGemmTraits44;
-  run_gemm<WmmaGemmTraits44>(4096, 4096, 4096);
-  #endif
-
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits25;
+//  run_gemm<WmmaGemmTraits25>(512, 512, 512);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits26;
+//  run_gemm<WmmaGemmTraits26>(512, 512, 512);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits27;
+//  run_gemm<WmmaGemmTraits27>(512, 512, 512);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits28;
+//  run_gemm<WmmaGemmTraits28>(512, 512, 512);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits29;
+//  run_gemm<WmmaGemmTraits29>(768, 768, 768);
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits30;
+//  run_gemm<WmmaGemmTraits30>(768, 768, 768);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits31;
+//  run_gemm<WmmaGemmTraits31>(768, 768, 768);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits32;
+//  run_gemm<WmmaGemmTraits32>(768, 768, 768);
+//
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits33;
+//  run_gemm<WmmaGemmTraits33>(1024, 1024, 1024);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits34;
+//  run_gemm<WmmaGemmTraits34>(1024, 1024, 1024);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits35;
+//  run_gemm<WmmaGemmTraits35>(1024, 1024, 1024);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits36;
+//  run_gemm<WmmaGemmTraits36>(1024, 1024, 1024);
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits37;
+//  run_gemm<WmmaGemmTraits37>(2048, 2048, 2048);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits38;
+//  run_gemm<WmmaGemmTraits38>(2048, 2048, 2048);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits39;
+//  run_gemm<WmmaGemmTraits39>(2048, 2048, 2048);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits40;
+//  run_gemm<WmmaGemmTraits40>(2048, 2048, 2048);
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits41;
+//  run_gemm<WmmaGemmTraits41>(4096, 4096, 4096);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits42;
+//  run_gemm<WmmaGemmTraits42>(4096, 4096, 4096);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits43;
+//  run_gemm<WmmaGemmTraits43>(4096, 4096, 4096);
+//
+//  typedef cutlass::gemm::WmmaGemmTraits<cutlass::MatrixLayout::kRowMajor,
+//                                        cutlass::MatrixLayout::kColumnMajor,
+//                                        cutlass::Shape<64, 128, 128> >
+//      WmmaGemmTraits44;
+//  run_gemm<WmmaGemmTraits44>(4096, 4096, 4096);