[VTA] Bringing group convolution support (apache#4421)

* group conv operator support for VTA

* autotvm tuning script for group conv2d

* lint fix

* lint fix

* lint fix

* addressing comments

vta/python/vta/top/__init__.py

@@ -22,5 +22,6 @@
 from . import op
 from . import vta_conv2d
 from . import vta_conv2d_transpose
+from . import vta_group_conv2d
 from . import vta_dense
 from . import util

vta/python/vta/top/vta_group_conv2d.py

@@ -0,0 +1,199 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""Group conv2D operator declaration and schedule registration for VTA."""
+
+import numpy as np
+
+import tvm
+from tvm import autotvm
+import topi
+
+from ..environment import get_env
+
+@autotvm.register_topi_compute(topi.nn.group_conv2d_nchw, 'vta', 'direct')
+def packed_group_conv2d(cfg,
+                        data,
+                        kernel,
+                        strides,
+                        padding,
+                        dilation,
+                        group,
+                        out_dtype):
+    """ Packed group conv2d nchw function."""
+    assert dilation == (1, 1)
+
+    if padding[0]:
+        pad_data = topi.nn.pad(data, [0, 0, padding[0], padding[1], 0, 0], name="pad_data")
+    else:
+        pad_data = data
+    assert len(data.shape) == 6
+    assert len(kernel.shape) == 6
+    assert data.dtype == "int8", data.dtype
+    assert kernel.dtype == "int8", kernel.dtype
+    assert out_dtype == "int32", out_dtype
+
+    oheight = topi.util.get_const_int((pad_data.shape[2] - kernel.shape[2]) // strides[0] + 1)
+    owidth = topi.util.get_const_int((pad_data.shape[3] - kernel.shape[3]) // strides[1] + 1)
+    oshape = (data.shape[0], kernel.shape[0], oheight, owidth, data.shape[4], kernel.shape[4])
+
+    ishape = topi.util.get_const_tuple(data.shape)
+    kshape = topi.util.get_const_tuple(kernel.shape)
+    assert group * kshape[1] == ishape[1]
+    assert kshape[0] % group == 0
+    d_i = tvm.reduce_axis((0, kshape[2]), name='d_i')
+    d_j = tvm.reduce_axis((0, kshape[3]), name='d_j')
+    k_o = tvm.reduce_axis((0, kshape[1]), name='k_o')
+    k_i = tvm.reduce_axis((0, kshape[-1]), name='k_i')
+    hstride, wstride = strides
+    out = tvm.compute(
+        oshape,
+        lambda b_o, c_o, i, j, b_i, c_i: tvm.sum(
+            pad_data[b_o, c_o // (kshape[0] // group) * kshape[1] + k_o, i * hstride + d_i,
+                     j * wstride + d_j, b_i, k_i].astype(out_dtype) *
+            kernel[c_o, k_o, d_i, d_j, c_i, k_i].astype(out_dtype),
+            axis=[k_o, d_i, d_j, k_i]),
+        name="res", tag="packed_group_conv2d")
+
+    cfg.add_flop(2 * np.prod(topi.util.get_const_tuple(oshape)) *
+                 kshape[2] * kshape[3] * ishape[1] * kshape[-1])
+
+    return out
+
+
+@autotvm.register_topi_schedule(topi.generic.schedule_group_conv2d_nchw, 'vta', 'direct')
+def schedule_packed_group_conv2d(cfg, outs):
+    """ Schedule the packed conv2d.
+    """
+    assert len(outs) == 1
+    output = outs[0]
+    const_ops = []
+    ewise_inputs = []
+    ewise_ops = []
+    conv2d_res = []
+    assert output.dtype == "int8"
+    assert output.op.input_tensors[0].dtype == "int32"
+
+    def _traverse(op):
+        if topi.tag.is_broadcast(op.tag):
+            if not op.same_as(output.op):
+                if not op.axis:
+                    const_ops.append(op)
+                else:
+                    ewise_ops.append(op)
+            for tensor in op.input_tensors:
+                if isinstance(tensor.op, tvm.tensor.PlaceholderOp):
+                    ewise_inputs.append((op, tensor))
+                else:
+                    _traverse(tensor.op)
+        else:
+            assert op.tag == "packed_group_conv2d"
+            conv2d_res.append(op)
+
+    _traverse(output.op)
+    assert len(conv2d_res) == 1
+    conv2d_stage = conv2d_res[0].output(0)
+    s = tvm.create_schedule(output.op)
+
+    ##### space definition begin #####
+    b, c_o, x_i, x_j, _, _ = s[conv2d_stage].op.axis
+    c_i, _, _, _ = s[conv2d_stage].op.reduce_axis
+    cfg.define_split('tile_b', b, num_outputs=2)
+    cfg.define_split('tile_h', x_i, num_outputs=2)
+    cfg.define_split('tile_w', x_j, num_outputs=2)
+    cfg.define_split('tile_ci', c_i, num_outputs=2)
+    cfg.define_split('tile_co', c_o, num_outputs=2)
+    cfg.define_knob('oc_nthread', [1, 2])
+    cfg.define_knob('h_nthread', [1, 2])
+    ###### space definition end ######
+
+    data, kernel = conv2d_stage.op.input_tensors
+    if isinstance(data.op, tvm.tensor.ComputeOp) and "pad" in data.op.tag:
+        temp = data.op.input_tensors[0]
+        pad_data = data
+        data = temp
+    else:
+        pad_data = None
+
+    env = get_env()
+
+    # setup pad
+    if pad_data is not None:
+        cdata = pad_data
+        s[pad_data].set_scope(env.inp_scope)
+    else:
+        cdata = s.cache_read(data, env.inp_scope, [conv2d_stage])
+    ckernel = s.cache_read(kernel, env.wgt_scope, [conv2d_stage])
+    s[conv2d_stage].set_scope(env.acc_scope)
+
+    # cache read input
+    cache_read_ewise = []
+    for consumer, tensor in ewise_inputs:
+        cache_read_ewise.append(
+            s.cache_read(tensor, env.acc_scope, [consumer]))
+
+    # set ewise scope
+    for op in ewise_ops:
+        s[op].set_scope(env.acc_scope)
+        s[op].pragma(s[op].op.axis[0], env.alu)
+
+    for op in const_ops:
+        s[op].compute_inline()
+
+    # tile
+    x_bo, x_co, x_i, x_j, x_bi, x_ci = s[output].op.axis
+    x_co0, x_co1 = cfg['tile_co'].apply(s, output, x_co)
+    x_i0, x_i1 = cfg['tile_h'].apply(s, output, x_i)
+    x_j0, x_j1 = cfg['tile_w'].apply(s, output, x_j)
+    s[output].reorder(x_bo, x_i0, x_co0, x_j0, x_co1, x_i1, x_j1, x_bi, x_ci)
+    store_pt = x_j0
+
+    # set all compute scopes
+    s[conv2d_stage].compute_at(s[output], store_pt)
+    for op in ewise_ops:
+        s[op].compute_at(s[output], store_pt)
+
+    for tensor in cache_read_ewise:
+        s[tensor].compute_at(s[output], store_pt)
+        s[tensor].pragma(s[tensor].op.axis[0], env.dma_copy)
+
+    # virtual threading along output channel axes
+    if cfg['oc_nthread'].val > 1:
+        _, v_t = s[output].split(x_co0, factor=cfg['oc_nthread'].val)
+        s[output].reorder(v_t, x_bo)
+        s[output].bind(v_t, tvm.thread_axis("cthread"))
+
+    # virtual threading along spatial rows
+    if cfg['h_nthread'].val > 1:
+        _, v_t = s[output].split(x_i0, factor=cfg['h_nthread'].val)
+        s[output].reorder(v_t, x_bo)
+        s[output].bind(v_t, tvm.thread_axis("cthread"))
+
+    x_bo, x_co, x_i, x_j, x_bi, x_ci = s[conv2d_stage].op.axis
+    k_o, d_i, d_j, k_i = s[conv2d_stage].op.reduce_axis
+    s[conv2d_stage].reorder(x_bo, k_o, x_j, d_j, d_i, x_co, x_i, x_bi, x_ci, k_i)
+
+    k_o, _ = cfg['tile_ci'].apply(s, conv2d_stage, k_o)
+    s[cdata].compute_at(s[conv2d_stage], k_o)
+    s[ckernel].compute_at(s[conv2d_stage], k_o)
+
+    # Use VTA instructions
+    s[cdata].pragma(s[cdata].op.axis[0], env.dma_copy)
+    s[ckernel].pragma(s[ckernel].op.axis[0], env.dma_copy)
+    s[conv2d_stage].tensorize(x_bi, env.gemm)
+    s[output].pragma(x_co1, env.dma_copy)
+
+    return s

vta/scripts/tune_group_conv2d.py

@@ -0,0 +1,155 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+"""Tuning a single group conv2d operator"""
+
+from collections import namedtuple
+import logging
+import os
+
+import tvm
+from tvm import autotvm
+from tvm.contrib.util import get_lower_ir
+import topi
+import vta
+import vta.testing
+
+env = vta.get_env()
+
+Workload = namedtuple("GroupConv2DWorkload",
+                      ['batch', 'height', 'width', 'in_filter', 'out_filter', 'groups',
+                       'hkernel', 'wkernel', 'hpad', 'wpad', 'hstride', 'wstride'])
+
+# Mobilenet (grouped variant) workloads
+mobilenet_wkls = [
+    ('mobilenet.D1', Workload(env.BATCH, 112, 112,   32,   32,  2, 3, 3, 1, 1, 1, 1)),
+    ('mobilenet.D2', Workload(env.BATCH, 112, 112,   64,   64,  4, 3, 3, 1, 1, 2, 2)),
+    ('mobilenet.D3', Workload(env.BATCH,  56,  56,  128,  128,  8, 3, 3, 1, 1, 1, 1)),
+    ('mobilenet.D4', Workload(env.BATCH,  56,  56,  128,  128,  8, 3, 3, 1, 1, 2, 2)),
+    ('mobilenet.D5', Workload(env.BATCH,  28,  28,  256,  256, 16, 3, 3, 1, 1, 1, 1)),
+    ('mobilenet.D6', Workload(env.BATCH,  28,  28,  256,  256, 16, 3, 3, 1, 1, 2, 2)),
+    ('mobilenet.D7', Workload(env.BATCH,  14,  14,  512,  512, 32, 3, 3, 1, 1, 1, 1)),
+    ('mobilenet.D8', Workload(env.BATCH,  14,  14,  512,  512, 32, 3, 3, 1, 1, 2, 2)),
+    ('mobilenet.D9', Workload(env.BATCH,   7,  7,  1024, 1024, 64, 3, 3, 1, 1, 1, 1)),
+]
+
+@tvm.tag_scope(tag=topi.tag.ELEMWISE)
+def my_clip(x, a_min, a_max):
+    """Unlike topi's current clip, put min and max into two stages."""
+    const_min = tvm.const(a_min, x.dtype)
+    const_max = tvm.const(a_max, x.dtype)
+    x = tvm.compute(x.shape, lambda *i: tvm.min(x(*i), const_max), name="clipA")
+    x = tvm.compute(x.shape, lambda *i: tvm.max(x(*i), const_min), name="clipB")
+    return x
+
+def group_conv2d(N, CI, H, W, CO, KH, KW, strides, padding, dilation, group):
+
+    CI_G = CI // groups
+    data_shape = (N//env.BATCH, CI//env.BLOCK_IN, H, W, env.BATCH, env.BLOCK_IN)
+    kernel_shape = (CO//env.BLOCK_OUT, CI_G//env.BLOCK_IN, KH, KW, env.BLOCK_OUT, env.BLOCK_IN)
+    bias_shape = (N//env.BATCH, CO//env.BLOCK_OUT, 1, 1, env.BATCH, env.BLOCK_OUT)
+
+    data = tvm.placeholder(data_shape, name="data", dtype=env.inp_dtype)
+    kernel = tvm.placeholder(kernel_shape, name="kernel", dtype=env.wgt_dtype)
+    bias = tvm.placeholder(bias_shape, name="bias", dtype=env.acc_dtype)
+
+    with tvm.target.vta():
+        res = topi.nn.group_conv2d_nchw(
+            data,
+            kernel,
+            strides,
+            padding,
+            dilation,
+            groups,
+            env.acc_dtype)
+        res = topi.right_shift(res, env.WGT_WIDTH)
+        res = topi.add(res, bias)
+        res = my_clip(res, 0, (1 << env.OUT_WIDTH - 1) - 1)
+        res = topi.cast(res, env.out_dtype)
+
+    if tvm.target.current_target().device_name == 'vta':
+        s = topi.generic.schedule_group_conv2d_nchw([res])
+    else:
+        s = tvm.create_schedule([res.op])
+
+    return s, [data, kernel, bias, res]
+
+if __name__ == '__main__':
+
+    # Logging config (for printing tuning log to the screen)
+    logging.basicConfig()
+
+    # Tuning log files
+    log_file = "%s.group_conv2d.log" % (env.TARGET)
+    # create tmp log file
+    tmp_log_file = log_file + ".tmp"
+    if os.path.exists(log_file):
+        os.remove(log_file)
+
+    # Get tracker info from env
+    tracker_host = os.environ.get("TVM_TRACKER_HOST", None)
+    tracker_port = os.environ.get("TVM_TRACKER_PORT", None)
+    if not tracker_host or not tracker_port:
+        print("Set your AutoTVM tracker node host and port variables to run the autotuner")
+        exit()
+
+    for idx, (wl_name, wl) in enumerate(mobilenet_wkls):
+        prefix = "[Task %2d/%2d] " % (idx, len(mobilenet_wkls))
+
+        # Read in workload parameters
+        N = wl.batch
+        CI = wl.in_filter
+        H = wl.height
+        W = wl.width
+        CO = wl.out_filter
+        KH = wl.hkernel
+        KW = wl.wkernel
+        strides = (wl.hstride, wl.wstride)
+        padding = (wl.hpad, wl.wpad)
+        dilation = (1, 1)
+        groups = wl.groups
+
+        # Create task
+        task = autotvm.task.create(
+                group_conv2d,
+                args=(N, CI, H, W, CO, KH, KW, strides, padding, dilation, groups),
+                target=tvm.target.vta(),
+                target_host=env.target_host,
+                template_key='direct')
+        print(task.config_space)
+
+        # Tune
+        measure_option = autotvm.measure_option(
+                builder=autotvm.LocalBuilder(),
+                runner=autotvm.RPCRunner(
+                    env.TARGET, host=tracker_host, port=int(tracker_port),
+                    number=5, timeout=60,
+                    check_correctness=True))
+
+        # Run Tuner
+        tuner = autotvm.tuner.RandomTuner(task)
+        tuner.tune(
+            n_trial=len(task.config_space),
+            early_stopping=None,
+            measure_option=measure_option,
+            callbacks=[
+                    autotvm.callback.progress_bar(len(task.config_space), prefix=prefix),
+                    autotvm.callback.log_to_file(tmp_log_file)])
+
+    # Pick best records to a cache file
+    autotvm.record.pick_best(tmp_log_file, log_file)
+    os.remove(tmp_log_file)