diff --git a/examples/python/tensorflow/dot.cpp b/examples/python/tensorflow/dot.cpp
index c31d8745c..183b3f492 100644
--- a/examples/python/tensorflow/dot.cpp
+++ b/examples/python/tensorflow/dot.cpp
@@ -23,7 +23,7 @@ const char* src =
 R"(
 const tunable int32 TM = {64, 128};
 const tunable int32 TN = {64, 128};
-const tunable int32 TK = {16};
+const tunable int32 TK = {32};
 const tunable int32 GZ = {1};
 
 void matmul(restrict read_only fp16 *A, restrict read_only fp16 *B,
@@ -117,10 +117,10 @@ class BlockSparseGemmOp : public OpKernel {
       return  2.*M*N*K / ts * 1e-3;
     };
     // just-in-time compile source-code
-    jit.autotune("matmul", src, benchmark);
+//    jit.autotune("matmul", src, benchmark);
 //    jit.add_module("matmul", src, {4, 2, 8, 4, 2, 32, 1, 4, 1, 1, 8, 8, 8, 1});
-//    jit.add_module("matmul", src, {32, 2, 128, 32, 2, 128, 2, 2, 2, 2, 4, 8, 4, 1});
     jit.add_module("matmul", src, {16, 4, 128, 16, 4, 128, 2, 2, 2, 2, 8, 32, 8, 1});
+//    jit.add_module("matmul", src, {8, 8, 128, 16, 8, 128, 2, 2, 2, 2, 16, 32, 8, 1 });
     triton::driver::kernel* kernel = jit.get_function("matmul");
     triton::jit::launch_information info = jit.get_launch_info("matmul");
     std::cout << benchmark(kernel, info) << std::endl;;
diff --git a/lib/codegen/selection.cpp b/lib/codegen/selection.cpp
index e265a0fc7..5c1d67bf3 100644
--- a/lib/codegen/selection.cpp
+++ b/lib/codegen/selection.cpp
@@ -501,8 +501,8 @@ void selection::init_axes(ir::value *v, IRBuilder<> &builder, Value *u_thread_id
     unsigned num_rep_0 = shapes[0]->get_value() / hmma_bts_0;
     unsigned num_rep_1 = shapes[1]->get_value() / hmma_bts_1;
     // size of each pack (interleaving)
-    pack_size_0_ = std::min<unsigned>(num_rep_0, 2);
-    pack_size_1_ = std::min<unsigned>(num_rep_1, 2);
+    pack_size_0_ = std::min<unsigned>(num_rep_0, 1);
+    pack_size_1_ = std::min<unsigned>(num_rep_1, 1);
     // number of packs (interleaving)
     num_packs_0_ = num_rep_0 / pack_size_0_;
     num_packs_1_ = num_rep_1 / pack_size_1_;
@@ -531,9 +531,14 @@ void selection::init_axes(ir::value *v, IRBuilder<> &builder, Value *u_thread_id
     // a offset
     offset_a_i_ = builder.CreateAdd(warp_offset_i, builder.CreateAdd(pair_a_off, in_pair_off_a));
     offset_a_k_ = builder.CreateAnd(u_thread_id, _3);
-    // b offsets
+//    // b offsets
     offset_b_j_ = builder.CreateAdd(warp_offset_j, builder.CreateAdd(pair_b_off, in_pair_off_b));
     offset_b_k_ = builder.CreateAnd(u_thread_id, _3);
+//    offset_a_i_ = builder.getInt32(0);
+//    offset_a_k_ = builder.getInt32(0);
+//    offset_b_j_ = builder.getInt32(0);
+//    offset_b_k_ = builder.getInt32(0);
+
     // c offsets
     Value *offset_c_i = builder.CreateAdd(builder.CreateAnd(u_thread_id, _1), offset_a_i_);
     Value *offset_c_j = builder.CreateAdd(builder.CreateAnd(u_thread_id, _2),
diff --git a/lib/codegen/shmem_allocation.cpp b/lib/codegen/shmem_allocation.cpp
index 6ef2101c2..aa7aada71 100644
--- a/lib/codegen/shmem_allocation.cpp
+++ b/lib/codegen/shmem_allocation.cpp
@@ -23,10 +23,11 @@ unsigned shmem_allocation::is_ld_padded(ir::value *x) {
   for(ir::user* user: x->get_users())
     if(dynamic_cast<ir::dot_inst*>(user))
     if(params_->get_fragment(user, 0) == tune::HMMA_FRAGMENT_C){
-      if(x == user->get_operand(0))
-        return 4;
+      if(x == user->get_operand(0)){
+        return 16;
+      }
       else
-        return 4;
+        return 16;
     }
   return 0;
 }
diff --git a/lib/driver/module.cpp b/lib/driver/module.cpp
index 4ff863666..c3139ece6 100755
--- a/lib/driver/module.cpp
+++ b/lib/driver/module.cpp
@@ -255,7 +255,7 @@ std::string cu_module::compile_llvm_module(llvm::Module* module) {
 cu_module::cu_module(driver::context * context, llvm::Module* ll_module): cu_module(context, compile_llvm_module(ll_module)) { }
 
 cu_module::cu_module(driver::context * context, std::string const & source) : module(context, CUmodule(), true), source_(source){
-//  std::cout << source << std::endl;
+  std::cout << source << std::endl;
   cu_context::context_switcher ctx_switch(*context);
   // JIT compile source-code
   CUjit_option opt[] = {CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES, CU_JIT_ERROR_LOG_BUFFER};