[DOC] Improve "Getting Started with TVM" tutorials and fix warnings (a…

…pache#8221)

* improve src/README.md

* fix intro

* fix more warnings

* improve docs

* update

* update

* update

* update overview image

docs/api/python/graph_executor.rst

@@ -16,6 +16,6 @@
     under the License.
 
 tvm.contrib.graph_executor
--------------------------
+--------------------------
 .. automodule:: tvm.contrib.graph_executor
     :members:

docs/conf.py

@@ -226,10 +226,10 @@ def git_describe_version(original_version):
         "introduction.py",
         "install.py",
         "tvmc_command_line_driver.py",
-        "auto_tuning_with_python.py",
+        "autotvm_relay_x86.py",
         "tensor_expr_get_started.py",
-        "autotvm_matmul.py",
-        "autoschedule_matmul.py",
+        "autotvm_matmul_x86.py",
+        "auto_scheduler_matmul_x86.py",
         "cross_compilation_and_rpc.py",
         "relay_quick_start.py",
     ],
@@ -246,7 +246,7 @@ def git_describe_version(original_version):
     ],
     "language": [
         "schedule_primitives.py",
-        "reduciton.py",
+        "reduction.py",
         "intrin_math.py",
         "scan.py",
         "extern_op.py",

docs/deploy/bnns.rst

@@ -175,7 +175,8 @@ Operator support
 | nn.bias_add            | Supported by BNNS integration only as a bias part of nn.conv2d or nn.dense   |
 |                        | fusion                                                                       |
 +------------------------+------------------------------------------------------------------------------+
-| add                    | Supported by BNNS integration only as a bias part of nn.conv2d or nn.dense fusion |
+| add                    | Supported by BNNS integration only as a bias part of nn.conv2d or nn.dense   |
+|                        | fusion                                                                       |
 +------------------------+------------------------------------------------------------------------------+
 | nn.relu                | Supported by BNNS integration only as a part of nn.conv2d or nn.dense fusion |
 +------------------------+------------------------------------------------------------------------------+

docs/dev/device_target_interactions.rst

@@ -18,7 +18,7 @@
 .. _tvm-target-specific-overview:
 
 Device/Target Interactions
---------------------------
+==========================
 
 This documented is intended for developers interested in understanding
 how the TVM framework interacts with specific device APIs, or who

docs/dev/index.rst

@@ -29,10 +29,6 @@ This page is organized as follows:
   The sections after are specific guides focused on each logical component, organized
   by the component's name.
 
-- The `Device/Target Interactions`_ section describes how TVM
-  interacts with each supported physical device and code-generation
-  target.
-
 - Feel free to also check out the :ref:`dev-how-to` for useful development tips.
 
 This guide provides a few complementary views of the architecture.
@@ -245,12 +241,13 @@ for learning-based optimizations.
 
 
 .. toctree::
-   :maxdepth: 2
+   :maxdepth: 1
 
    runtime
    debugger
    virtual_machine
    introduction_to_module_serialization
+   device_target_interactions
 
 tvm/node
 --------
@@ -318,6 +315,11 @@ It also provides a common `Target` class that describes the target.
 The compilation pipeline can be customized according to the target by querying the attribute information
 in the target and builtin information registered to each target id(cuda, opencl).
 
+.. toctree::
+   :maxdepth: 1
+
+   device_target_interactions
+
 tvm/tir
 -------
 

docs/index.rst

@@ -44,8 +44,8 @@ For Developers
    contribute/index
    deploy/index
    dev/how_to
-   microtvm/index
    errors
+   faq
 
 .. toctree::
    :maxdepth: 1
@@ -76,8 +76,8 @@ For Developers
    :hidden:
    :caption: MISC
 
+   microtvm/index
    vta/index
-   faq
 
 
 Index

python/tvm/relay/op/nn/nn.py

@@ -2236,6 +2236,7 @@ def sparse_add(dense_mat, sparse_mat):
     Examples
     -------
     .. code-block:: python
+
         dense_data = [[ 3.,   4.,   4. ]
                       [ 4.,  2.,  5. ]]
         sparse_data = [4., 8.]

python/tvm/relay/op/transform.py

@@ -1404,6 +1404,7 @@ def sparse_fill_empty_rows(sparse_indices, sparse_values, dense_shape, default_v
     Examples
     -------
     .. code-block:: python
+
         sparse_indices = [[0, 1],
                          [0, 3],
                          [2, 0],
@@ -1425,7 +1426,6 @@ def sparse_fill_empty_rows(sparse_indices, sparse_values, dense_shape, default_v
                              [4, 0]]
         empty_row_indicator = [False, True, False, False, True]
         new_sparse_values = [1, 2, 10, 3, 4, 10]
-
     """
     new_sparse_indices, new_sparse_values, empty_row_indicator = TupleWrapper(
         _make.sparse_fill_empty_rows(sparse_indices, sparse_values, dense_shape, default_value), 3
@@ -1457,6 +1457,7 @@ def sparse_reshape(sparse_indices, prev_shape, new_shape):
     Examples
     --------
     .. code-block:: python
+
         sparse_indices = [[0, 0, 0],
                             [0, 0, 1],
                             [0, 1, 0],
@@ -1508,6 +1509,7 @@ def segment_sum(data, segment_ids, num_segments=None):
     Examples
     --------
     .. code-block:: python
+
         data = [[1, 2, 3, 4],
                 [4, -3, 2, -1],
                 [5, 6, 7, 8]]
@@ -1578,6 +1580,7 @@ def cumsum(data, axis=None, dtype=None, exclusive=None):
     Examples
     --------
     .. code-block:: python
+
         a = [[1,2,3], [4,5,6]]
 
         cumsum(a)  # if axis is not provided, cumsum is done over the flattened input.
@@ -1633,6 +1636,7 @@ def cumprod(data, axis=None, dtype=None, exclusive=None):
     Examples
     --------
     .. code-block:: python
+
         a = [[1,2,3], [4,5,6]]
 
         cumprod(a)  # if axis is not provided, cumprod is done over the flattened input.
@@ -1693,6 +1697,7 @@ def unique(data, is_sorted=True, return_counts=False):
     Examples
     --------
     .. code-block:: python
+
         [output, indices, num_unique] = unique([4, 5, 1, 2, 3, 3, 4, 5], False, False)
         output         =  [4, 5, 1, 2, 3, ?, ?, ?]
         indices        =  [0, 1, 2, 3, 4, 4, 0, 1]

python/tvm/topi/cuda/sparse_reshape.py

@@ -48,6 +48,7 @@ def sparse_reshape(
     Examples
     --------
     .. code-block:: python
+
         sparse_indices = [[0, 0, 0],
                             [0, 0, 1],
                             [0, 1, 0],

python/tvm/topi/cuda/unique.py

@@ -317,6 +317,7 @@ def unique(data, is_sorted=True, return_counts=False):
     Examples
     --------
     .. code-block:: python
+
         [output, indices, num_unique] = unique([4, 5, 1, 2, 3, 3, 4, 5], False, False)
         output          =  [4, 5, 1, 2, 3, ?, ?, ?]
         indices         =  [0, 1, 2, 3, 4, ?, ?, ?]

python/tvm/topi/sparse_reshape.py

@@ -45,6 +45,7 @@ def sparse_reshape(
     Examples
     --------
     .. code-block:: python
+
         sparse_indices = [[0, 0, 0],
                             [0, 0, 1],
                             [0, 1, 0],

python/tvm/topi/unique.py

@@ -243,6 +243,7 @@ def unique(data, is_sorted=True, return_counts=False):
     Examples
     --------
     .. code-block:: python
+
         [output, indices, num_unique] = unique([4, 5, 1, 2, 3, 3, 4, 5], False, False)
         output          =  [4, 5, 1, 2, 3, ?, ?, ?]
         indices         =  [0, 1, 2, 3, 4, ?, ?, ?]

src/README.md

@@ -21,14 +21,17 @@ Header files in include are public APIs that share across modules.
 There can be internal header files within each module that sit in src.
 
 ## Modules
-- support: Internal support utilities.
-- runtime: Minimum runtime related codes.
-- node: base infra for IR/AST nodes that is dialect independent.
-- ir: Common IR infrastructure.
-- tir: Tensor-level IR.
-- te: tensor expression DSL
 - arith: Arithmetic expression and set simplification.
-- relay: Relay IR, high-level optimization.
-- autotvm: The auto-tuning module.
+- auto\_scheduler: The template-free auto-tuning module.
+- autotvm: The template-based auto-tuning module.
 - contrib: Contrib extension libraries.
 - driver: Compilation driver APIs.
+- ir: Common IR infrastructure.
+- node: The base infra for IR/AST nodes that is dialect independent.
+- relay: Relay IR, high-level optimizations.
+- runtime: Minimum runtime related codes.
+- support: Internal support utilities.
+- target: Hardwaer target.
+- tir: Tensor IR, low-level optimizations.
+- te: Tensor expression DSL.
+- topi: Tensor Operator Inventory.

tutorials/auto_scheduler/tune_network_arm.py

@@ -437,7 +437,7 @@ def tune_and_evaluate():
 #    in function :code:`run_tuning`. Say,
 #    :code:`tuner = auto_scheduler.TaskScheduler(tasks, task_weights, load_log_file=log_file)`
 # 4. If you have multiple target CPUs, you can use all of them for measurements to
-#    parallelize the measurements. Check this :ref:`section <tutorials-autotvm-rpc-tracker>`
+#    parallelize the measurements. Check this :ref:`section <tutorials-autotvm-scale-up-rpc-tracker>`
 #    to learn how to use the RPC Tracker and RPC Server.
 #    To use the RPC Tracker in auto-scheduler, replace the runner in :code:`TuningOptions`
 #    with :any:`auto_scheduler.RPCRunner`.

tutorials/autotvm/tune_conv2d_cuda.py

@@ -77,7 +77,7 @@
 # to tune other operators such as depthwise convolution and gemm.
 # In order to fully understand this template, you should be familiar with
 # the schedule primitives and auto tuning API. You can refer to the above
-# tutorials and :doc:`autotvm tutorial <tune_simple_template>`
+# tutorials and :ref:`autotvm tutorial <tutorial-autotvm-matmul-x86>`
 #
 # It is worth noting that the search space for a conv2d operator
 # can be very large (at the level of 10^9 for some input shapes)

tutorials/get_started/tune_matmul_x86.py → tutorials/get_started/auto_scheduler_matmul_x86.py

@@ -23,7 +23,7 @@
 In this tutorial, we will show how TVM's Auto Scheduling feature can find
 optimal schedules without the need for writing a custom template.
 
-Different from the template-based :ref:`<autotvm_matmul>` which relies on
+Different from the template-based :doc:`AutoTVM <autotvm_matmul_x86>` which relies on
 manual templates to define the search space, the auto-scheduler does not
 require any templates.  Users only need to write the computation declaration
 without any schedule commands or templates.  The auto-scheduler can

tutorials/get_started/autotvm_matmul.py → tutorials/get_started/autotvm_matmul_x86.py

@@ -15,17 +15,18 @@
 # specific language governing permissions and limitations
 # under the License.
 """
-Optimizing Operators with Templates and AutoTVM
-===============================================
+.. _tutorial-autotvm-matmul-x86:
+
+Optimizing Operators with Schedule Templates and AutoTVM
+========================================================
 **Authors**:
 `Lianmin Zheng <https://github.com/merrymercy>`_,
 `Chris Hoge <https://github.com/hogepodge>`_
 
-In this tutorial, we will now show how the TVM Template Extension (TE) language
-can be used to write scheduling templates that can be searched by AutoTVM to
-find optimal configurations of scheduling variables. This process is called
-Auto-Tuning, and builds on TE to help automate the process of optimizing
-operations.
+In this tutorial, we show how the TVM Tensor Expression (TE) language
+can be used to write schedule templates that can be searched by AutoTVM to
+find the optimal schedule. This process is called Auto-Tuning, which helps
+automate the process of optimizing tensor computation.
 
 This tutorial builds on the previous `tutorial on how to write a matrix
 multiplication using TE <tensor_expr_get_started>`.
@@ -371,6 +372,6 @@ def matmul(N, L, M, dtype):
 # To gain a deeper understanding of how this works, we recommend expanding on
 # this example by adding new search parameters to the schedule based on
 # schedule operations demonstated in the `Getting Started With Tensor
-# Expressions <tensor_expr_get_started>_` tutorial In the upcoming sections, we
+# Expressions <tensor_expr_get_started>_` tutorial. In the upcoming sections, we
 # will demonstate the AutoScheduler, a method for TVM to optimize common
 # operators without the need for the user to provide a user-defined template.

tutorials/get_started/auto_tuning_with_python.py → tutorials/get_started/autotvm_relay_x86.py

@@ -15,8 +15,8 @@
 # specific language governing permissions and limitations
 # under the License.
 """
-Compiling and Optimizing a Model with the Python AutoScheduler
-==============================================================
+Compiling and Optimizing a Model with the Python Interface (AutoTVM)
+====================================================================
 **Author**:
 `Chris Hoge <https://github.com/hogepodge>`_
 
@@ -302,6 +302,7 @@
     repeat=repeat,
     timeout=timeout,
     min_repeat_ms=min_repeat_ms,
+    enable_cpu_cache_flush=True,
 )
 
 # Create a simple structure for holding tuning options. We use an XGBoost

tutorials/get_started/install.py

@@ -23,8 +23,8 @@
 
 Depending on your needs and your working environment, there are a few different
 methods for installing TVM. These include:
-    * Installing from source 
-    * Installing from third-party binary package.
+* Installing from source
+* Installing from third-party binary package.
 """
 
 ################################################################################