testing GEMM

examples/python/tensorflow/run.py

@@ -9,7 +9,7 @@ library_dir = os.path.dirname(os.path.realpath(__file__))
 module = tf.load_op_library(os.path.join(library_dir, 'libtf_blocksparse.so'))
 
 def run_dot():
-    M, N, K = 8192, 8192, 8192
+    M, N, K = 128, 128, 128
     a = tf.placeholder(tf.float16, shape=[M, K])
     b = tf.placeholder(tf.float16, shape=[N, K])
     # c = tf.matmul(a, b, transpose_a=True)
@@ -23,9 +23,11 @@ def run_dot():
     result = sess.run([c], feed_dict = {a: ha,
                                         b: hb})[0]
     # Test
-    #hresult = np.dot(ha.T, hb).T
-    #dif = np.abs(result - hresult)
-    #print("dif: %f" % np.max(dif))
+    hresult = np.dot(ha.T, hb).T
+    dif = np.abs(result - hresult)
+    print(hresult)
+    print(result)
+    print("dif: %f" % np.max(dif))
 
 def run_conv():
     B, C, H, W = 16, 32, 32, 32
@@ -128,6 +130,6 @@ def run_batchnorm():
     print(np.max(np.abs(dg_t - dg_n)))
     print(np.max(np.abs(db_t - db_n)))
 
-#run_dot()
-run_shift()
+run_dot()
+#run_shift()
 #run_batchnorm()