preparing the field for tensor cores transposes

examples/python/tensorflow/dot.cpp

@@ -49,7 +49,7 @@ class DotOp : public OpKernel {
     triton::driver::cu_buffer db(ctx, (CUdeviceptr)b.flat<Eigen::half>().data(), false);
     triton::driver::cu_buffer dc(ctx, (CUdeviceptr)c->flat<float>().data(), false);
     // template
-    triton::dnn::gemm dot(M, N, K, false, true, "fp16", "fp16", 4, 4);
+    triton::dnn::gemm dot(M, N, K, false, false, "fp16", "fp16", 4, 4);
     dot.enqueue(stream, {&da, &db, &dc});
   }
 

examples/python/tensorflow/run.py

@@ -23,7 +23,7 @@ def run_dot():
     result = sess.run([c], feed_dict = {a: ha,
                                         b: hb})[0]
     # Test
-    hresult = np.dot(ha.T, hb).T
+    hresult = np.dot(ha.T, hb.T).T
     dif = np.abs(result - hresult)
     print(hresult)
     print(result)