# Copyright (c) 2020 Graphcore Ltd. All rights reserved. """Tutorial code to play around with graph recompilation and executable loading Parameters to play around with are CACHING, NOMULTISESSION, PLACEHOLDER, and SAMEBATCH. Some comments in the document refer to the underlying tutorial in the documentation portal. The code will print out what the expected behaviour should look like. """ import os import numpy as np import tensorflow as tf from tensorflow.python import ipu from tensorflow.python.ipu.scopes import ipu_scope # Consideration 0: Environment setup CACHING = True # Cache compiled graph. The folder is tmp_tutorial. # Consideration 1: Sessions NOMULTISESSION = True # Avoid using different sessions. # Consideration 2, 4, 5: Graphs, Weights, Constants # Use a placeholder that is handed over to the graph instead of a hard coded # hyperparameter that might change between executions. PLACEHOLDER = True # Consideration 3: Batch size SAMEBATCH = True # Change the batch size between executions. # Consideration 0: Environment setup if "TF_POPLAR_FLAGS" in os.environ and not CACHING: os.environ["TF_POPLAR_FLAGS"] = "" else: os.environ["TF_POPLAR_FLAGS"] = "--executable_cache_path=tmp_tutorial" if "POPLAR_LOG_LEVEL" not in os.environ or \ os.environ["POPLAR_LOG_LEVEL"] != "INFO": print("Setting POPLAR_LOG_LEVEL to INFO for graph compilation information.") os.environ["POPLAR_LOG_LEVEL"] = "INFO" # Consideration 6 os.environ["XLA_FLAGS"] = "--xla_dump_to=tmp_xla_{} ".format( np.random.randint(2, 101)) os.environ["XLA_FLAGS"] += " --xla_dump_hlo_pass_re=forward-allocation " os.environ["XLA_FLAGS"] += " --xla_hlo_graph_sharding_color " os.environ["XLA_FLAGS"] += " --xla_dump_hlo_as_text " # Configure arguments for targeting the IPU cfg = ipu.utils.create_ipu_config() cfg = ipu.utils.auto_select_ipus(cfg, 1) ipu.utils.configure_ipu_system(cfg) with tf.device("cpu"): pa = tf.placeholder(np.float32, [None, 2], name="a") pb = tf.placeholder(np.float32, [None, 2], name="b") pc = tf.placeholder(np.float32, [None, 2], name="c") if PLACEHOLDER: mult = tf.placeholder(np.float32, [], name="multiplier") else: mult = np.random.uniform(0, 1) def basic_graph(pa, pb, pc): # Do basic addition with tensors o1 = pa + pb o2 = pa + pc simple_graph_output = mult * (o1 + o2) return simple_graph_output with ipu_scope("/device:IPU:0"): comp_graph = basic_graph(pa, pb, pc) print("\nWarm up & Caching Test: ") print("No compilation after first execution expected but executable load. \n") with tf.Session() as sess1, tf.Session() as sess2: # Run the graph through the session feeding it an arbitrary dictionary if PLACEHOLDER: result0 = sess1.run(comp_graph, feed_dict={ pa: [[1., 1.]], pb: [[0., 1.]], pc: [[1., 5.]], mult: 10.0 }) else: result0 = sess1.run(comp_graph, feed_dict={ pa: [[1., 1.]], pb: [[0., 1.]], pc: [[1., 5.]], }) # Consideration 6 os.environ["XLA_FLAGS"] = "--xla_dump_to=tmp_xla_{} ".format( np.random.randint(101, 201)) os.environ["XLA_FLAGS"] += " --xla_dump_hlo_pass_re=forward-allocation " os.environ["XLA_FLAGS"] += " --xla_hlo_graph_sharding_color " os.environ["XLA_FLAGS"] += " --xla_dump_hlo_as_text " # Consideration 2, 4, 5: Graphs, Weights, Constants m = np.random.uniform(0, 1) if not PLACEHOLDER: mult = m with ipu_scope("/device:IPU:0"): comp_graph = basic_graph(pa, pb, pc) with tf.Session() as sess1, tf.Session() as sess2: print("\nPlaceholder test. ") print("No recompilation but executable switch should occur.\n") # Run the graph through the session feeding it an arbitrary dictionary if PLACEHOLDER: result1 = sess1.run(comp_graph, feed_dict={ pa: [[1., 1.]], pb: [[0., 1.]], pc: [[1., 5.]], mult: m }) else: result1 = sess1.run(comp_graph, feed_dict={ pa: [[1., 1.]], pb: [[0., 1.]], pc: [[1., 5.]], }) # Consideration 6 os.environ["XLA_FLAGS"] = "--xla_dump_to=tmp_xla_{} ".format( np.random.randint(201, 301)) os.environ["XLA_FLAGS"] += " --xla_dump_hlo_pass_re=forward-allocation " os.environ["XLA_FLAGS"] += " --xla_hlo_graph_sharding_color " os.environ["XLA_FLAGS"] += " --xla_dump_hlo_as_text " # Consideration 1: Sessions if NOMULTISESSION: sess2 = sess1 else: print("Switching session.") print("\nSession Test.") print("No recompilation or executable switch should occur.\n") if PLACEHOLDER: result2 = sess2.run(comp_graph, feed_dict={ pa: [[1., 1.]], pb: [[0., 1.]], pc: [[1., 5.]], mult: m }) else: result2 = sess2.run(comp_graph, feed_dict={ pa: [[1., 1.]], pb: [[0., 1.]], pc: [[1., 5.]], }) # Consideration 6 os.environ["XLA_FLAGS"] = "--xla_dump_to=tmp_xla_{} ".format( np.random.randint(301, 401)) os.environ["XLA_FLAGS"] += " --xla_dump_hlo_pass_re=forward-allocation " os.environ["XLA_FLAGS"] += " --xla_hlo_graph_sharding_color " os.environ["XLA_FLAGS"] += " --xla_dump_hlo_as_text " # Consideration 3: Batch size if SAMEBATCH: bs = 1 else: bs = np.random.randint(2, 101) print("\nBatch Size Test with batch size %d." % bs) print("No recompilation or executable switch should occur.") print("Batch size should be the original 1.\n") if PLACEHOLDER: result3 = sess2.run(comp_graph, feed_dict={ pa: [[1., 1.]] * bs, pb: [[0., 1.]] * bs, pc: [[1., 5.]] * bs, mult: m }) else: result3 = sess2.run(comp_graph, feed_dict={ pa: [[1., 1.]] * bs, pb: [[0., 1.]] * bs, pc: [[1., 5.]] * bs, }) print("\nFirst two results should be different (different multiplier).\n") print("Caching/warm up test:\t", result0) print() print("Placeholder test: \t", result1) print() print("Session test: \t", result2) print() if bs > 1: print("Batch size test: \t", result3[:2], "...") else: print("Batch size test: \t", result3)