Tuning High Performance Convolution on NVIDIA GPUs

Author: Lianmin Zheng

This is an advanced tutorial for writing high performance tunable template for NVIDIA GPU. By running auto-tuner on this template, we can outperform the vendor provided library CuDNN in many cases.

Note that this tutorial will not run on Windows or recent versions of macOS. To get it to run, you will need to wrap the body of this tutorial in a if __name__ == "__main__": block.

Install dependencies

To use autotvm package in tvm, we need to install some extra dependencies. (change “3” to “2” if you use python2):

pip3 install --user psutil xgboost tornado

To make TVM run faster in tuning, it is recommended to use cython as FFI of tvm. In the root directory of tvm, execute

pip3 install --user cython
sudo make cython3

Now return to python code. Import packages.

import logging
import sys
import numpy as np

import tvm
from tvm import te, topi, testing
from tvm.topi.testing import conv2d_nchw_python

from tvm import autotvm

Step 1: Define the search space

There are plenty of useful schedule primitives in tvm. You can also find some tutorials that describe them in more details, such as (1). How to optimize convolution on GPU (2). Optimizing DepthwiseConv on NVIDIA GPU

However, their implementations are manually tuned for some special input shapes. In this section, we build a large enough space to cover the techniques used in these tutorials. Then we rely on the efficient auto-tuner to search through this space and pick some good configurations.

If you are familiar with writing cuda schedule, you can find the following template is very general. Actually this template can be easily modified 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 autotvm tutorial

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)

@autotvm.template("tutorial/conv2d_no_batching")
def conv2d_no_batching(N, H, W, CO, CI, KH, KW, stride, padding):
    assert N == 1, "Only consider batch_size = 1 in this template"

    data = te.placeholder((N, CI, H, W), name="data")
    kernel = te.placeholder((CO, CI, KH, KW), name="kernel")
    conv = topi.nn.conv2d_nchw(data, kernel, stride, padding, dilation=1, out_dtype="float32")
    s = te.create_schedule([conv.op])

    ##### space definition begin #####
    n, f, y, x = s[conv].op.axis
    rc, ry, rx = s[conv].op.reduce_axis

    cfg = autotvm.get_config()
    cfg.define_split("tile_f", f, num_outputs=4)
    cfg.define_split("tile_y", y, num_outputs=4)
    cfg.define_split("tile_x", x, num_outputs=4)
    cfg.define_split("tile_rc", rc, num_outputs=3)
    cfg.define_split("tile_ry", ry, num_outputs=3)
    cfg.define_split("tile_rx", rx, num_outputs=3)
    cfg.define_knob("auto_unroll_max_step", [0, 512, 1500])
    cfg.define_knob("unroll_explicit", [0, 1])
    ##### space definition end #####

    # inline padding
    pad_data = s[conv].op.input_tensors[0]
    s[pad_data].compute_inline()
    data, raw_data = pad_data, data

    output = conv
    OL = s.cache_write(conv, "local")

    # create cache stage
    AA = s.cache_read(data, "shared", [OL])
    WW = s.cache_read(kernel, "shared", [OL])
    AL = s.cache_read(AA, "local", [OL])
    WL = s.cache_read(WW, "local", [OL])

    # tile and bind spatial axes
    n, f, y, x = s[output].op.axis
    bf, vf, tf, fi = cfg["tile_f"].apply(s, output, f)
    by, vy, ty, yi = cfg["tile_y"].apply(s, output, y)
    bx, vx, tx, xi = cfg["tile_x"].apply(s, output, x)
    kernel_scope = n  # this is the scope to attach global config inside this kernel

    s[output].bind(bf, te.thread_axis("blockIdx.z"))
    s[output].bind(by, te.thread_axis("blockIdx.y"))
    s[output].bind(bx, te.thread_axis("blockIdx.x"))
    s[output].bind(vf, te.thread_axis("vthread"))
    s[output].bind(vy, te.thread_axis("vthread"))
    s[output].bind(vx, te.thread_axis("vthread"))
    s[output].bind(tf, te.thread_axis("threadIdx.z"))
    s[output].bind(ty, te.thread_axis("threadIdx.y"))
    s[output].bind(tx, te.thread_axis("threadIdx.x"))
    s[output].reorder(n, bf, by, bx, vf, vy, vx, tf, ty, tx, fi, yi, xi)
    s[OL].compute_at(s[output], tx)

    # tile reduction axes
    n, f, y, x = s[OL].op.axis
    rc, ry, rx = s[OL].op.reduce_axis
    rco, rcm, rci = cfg["tile_rc"].apply(s, OL, rc)
    ryo, rym, ryi = cfg["tile_rx"].apply(s, OL, ry)
    rxo, rxm, rxi = cfg["tile_ry"].apply(s, OL, rx)
    s[OL].reorder(rco, ryo, rxo, rcm, rym, rxm, rci, ryi, rxi, n, f, y, x)

    s[AA].compute_at(s[OL], rxo)
    s[WW].compute_at(s[OL], rxo)
    s[AL].compute_at(s[OL], rxm)
    s[WL].compute_at(s[OL], rxm)

    # cooperative fetching
    for load in [AA, WW]:
        n, f, y, x = s[load].op.axis
        fused = s[load].fuse(n, f, y, x)
        tz, fused = s[load].split(fused, nparts=cfg["tile_f"].size[2])
        ty, fused = s[load].split(fused, nparts=cfg["tile_y"].size[2])
        tx, fused = s[load].split(fused, nparts=cfg["tile_x"].size[2])
        s[load].bind(tz, te.thread_axis("threadIdx.z"))
        s[load].bind(ty, te.thread_axis("threadIdx.y"))
        s[load].bind(tx, te.thread_axis("threadIdx.x"))

    # tune unroll
    s[output].pragma(kernel_scope, "auto_unroll_max_step", cfg["auto_unroll_max_step"].val)
    s[output].pragma(kernel_scope, "unroll_explicit", cfg["unroll_explicit"].val)

    return s, [raw_data, kernel, conv]

Step 2: Search through the space

We pick the last layer on resnet as test case. Since our space is very large, XGBoostTuner is most suitable for our case. Here we only do 20 trials for demonstration. In practice, making 1000 trials usually can find some good kernels for this template

# logging config (for printing tuning log to screen)
logging.getLogger("autotvm").setLevel(logging.DEBUG)
logging.getLogger("autotvm").addHandler(logging.StreamHandler(sys.stdout))

# the last layer in resnet
N, H, W, CO, CI, KH, KW, strides, padding = 1, 7, 7, 512, 512, 3, 3, (1, 1), (1, 1)
task = autotvm.task.create(
    "tutorial/conv2d_no_batching", args=(N, H, W, CO, CI, KH, KW, strides, padding), target="cuda"
)
print(task.config_space)

# Use local gpu, measure 10 times for every config to reduce variance
# The timeout of compiling a program is 10 seconds, the timeout for running is 4 seconds
measure_option = autotvm.measure_option(
    builder=autotvm.LocalBuilder(),
    runner=autotvm.LocalRunner(repeat=3, min_repeat_ms=100, timeout=4),
)

# Begin tuning, log records to file `conv2d.log`
# During tuning we will also try many invalid configs, so you are expected to
# see many error reports. As long as you can see non-zero GFLOPS, it is okay.
tuner = autotvm.tuner.XGBTuner(task)
tuner.tune(
    n_trial=20,
    measure_option=measure_option,
    callbacks=[autotvm.callback.log_to_file("conv2d.log")],
)

Out:

ConfigSpace (len=10454400, space_map=
   0 tile_f: Split(policy=factors, product=512, num_outputs=4) len=220
   1 tile_y: Split(policy=factors, product=7, num_outputs=4) len=4
   2 tile_x: Split(policy=factors, product=7, num_outputs=4) len=4
   3 tile_rc: Split(policy=factors, product=512, num_outputs=3) len=55
   4 tile_ry: Split(policy=factors, product=3, num_outputs=3) len=3
   5 tile_rx: Split(policy=factors, product=3, num_outputs=3) len=3
   6 auto_unroll_max_step: OtherOption([0, 512, 1500]) len=3
   7 unroll_explicit: OtherOption([0, 1]) len=2
)
Get devices for measurement successfully!
No: 1   GFLOPS: 456.68/456.68   result: MeasureResult(costs=(0.0005069250867924528,), error_no=0, all_cost=1.483231544494629, timestamp=1606247193.9099)        [('tile_f', [-1, 2, 64, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4881186
No: 2   GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.07772326469421387, timestamp=1606247191.9689252) [('tile_f', [-1, 8, 4, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 16, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,7106118
No: 3   GFLOPS: 220.25/456.68   result: MeasureResult(costs=(0.0010510892745098038,), error_no=0, all_cost=1.5199861526489258, timestamp=1606247195.1761348)    [('tile_f', [-1, 4, 32, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3605182
No: 4   GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.11786580085754395, timestamp=1606247191.9836805) [('tile_f', [-1, 8, 2, 32]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 4, 16]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9420813
No: 5   GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.053978919982910156, timestamp=1606247191.983764) [('tile_f', [-1, 8, 2, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 1, 128]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1721347
No: 6   GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.047043800354003906, timestamp=1606247191.9838579)        [('tile_f', [-1, 1, 64, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 32, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 1)],None,6444505
No: 7   GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.07104849815368652, timestamp=1606247191.9839547) [('tile_f', [-1, 4, 2, 2]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 16, 32]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 512), ('unroll_explicit', 1)],None,8675766
No: 8   GFLOPS: 2.18/456.68     result: MeasureResult(costs=(0.10620077,), error_no=0, all_cost=3.1617162227630615, timestamp=1606247198.2543473)       [('tile_f', [-1, 2, 4, 64]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2716108
No: 9   GFLOPS: 21.16/456.68    result: MeasureResult(costs=(0.010940765500000001,), error_no=0, all_cost=1.6032164096832275, timestamp=1606247199.146415)      [('tile_f', [-1, 1, 4, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,1263092
No: 10  GFLOPS: 273.42/456.68   result: MeasureResult(costs=(0.0008466825684210527,), error_no=0, all_cost=1.5583219528198242, timestamp=1606247200.2024171)    [('tile_f', [-1, 1, 32, 4]), ('tile_y', [-1, 1, 7, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 16, 1]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,8921130
No: 11  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.08771109580993652, timestamp=1606247193.050288)  [('tile_f', [-1, 32, 2, 1]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 64]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4036575
No: 12  GFLOPS: 149.26/456.68   result: MeasureResult(costs=(0.0015510021076923077,), error_no=0, all_cost=1.2667763233184814, timestamp=1606247201.3664002)    [('tile_f', [-1, 2, 32, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 1, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5036371
No: 13  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.0962529182434082, timestamp=1606247193.0504525)  [('tile_f', [-1, 4, 4, 8]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 64, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 0), ('unroll_explicit', 0)],None,252051
No: 14  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.06756925582885742, timestamp=1606247193.0505393) [('tile_f', [-1, 4, 64, 2]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 64, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5057676
No: 15  GFLOPS: 108.20/456.68   result: MeasureResult(costs=(0.0021395734680851063,), error_no=0, all_cost=1.2361633777618408, timestamp=1606247202.6363714)    [('tile_f', [-1, 1, 1, 4]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 1, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,3582580
No: 16  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.08173298835754395, timestamp=1606247193.050694)  [('tile_f', [-1, 4, 8, 16]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 1]), ('tile_rc', [-1, 128, 1]), ('tile_ry', [-1, 1, 3]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 1)],None,9705041
No: 17  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.031887054443359375, timestamp=1606247203.5122843)        [('tile_f', [-1, 1, 512, 1]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 16, 4]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4730274
No: 18  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.06038999557495117, timestamp=1606247203.6010818) [('tile_f', [-1, 1, 16, 2]), ('tile_y', [-1, 1, 1, 1]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 64]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,5003765
No: 19  GFLOPS: 15.67/456.68    result: MeasureResult(costs=(0.014774698,), error_no=0, all_cost=1.5089759826660156, timestamp=1606247205.8270934)      [('tile_f', [-1, 8, 64, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 1, 7]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4107668
No: 20  GFLOPS: 0.00/456.68     result: MeasureResult(costs=(InstantiationError('Traceback (most recent call last):\n  [bt] (4) /workspace/build/libtvm.so(TVMFuncCall+0x61) [0x7f9267ab0531]\n  [bt] (3) /workspace/build/libtvm.so(+0x6dbf77) [0x7f9266ed2f77]\n  [bt] (2) /workspace/build/libtvm.so(tvm::transform::SequentialNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x40e) [0x7f9266ed238e]\n  [bt] (1) /workspace/build/libtvm.so(tvm::tir::transform::PrimFuncPassNode::operator()(tvm::IRModule, tvm::transform::PassContext const&) const+0x55f) [0x7f92671561af]\n  [bt] (0) /workspace/build/libtvm.so(+0x12b5f3b) [0x7f9267aacf3b]\n  File "tvm/_ffi/_cython/./packed_func.pxi", line 56, in tvm._ffi._cy3.core.tvm_callback\n  File "/workspace/docs/../python/tvm/autotvm/measure/measure_methods.py", line 747, in verify_pass\n    raise InstantiationError("Skipped because of invalid gpu kernel")\ntvm.autotvm.task.space.InstantiationError: Skipped because of invalid gpu kernel',),), error_no=1, all_cost=0.06198382377624512, timestamp=1606247204.3836548) [('tile_f', [-1, 4, 8, 8]), ('tile_y', [-1, 7, 1, 1]), ('tile_x', [-1, 7, 1, 1]), ('tile_rc', [-1, 64, 8]), ('tile_ry', [-1, 1, 1]), ('tile_rx', [-1, 3, 1]), ('auto_unroll_max_step', 512), ('unroll_explicit', 0)],None,2440616

Finally we can inspect the best config from log file, check correctness, and measure running time.

# inspect the best config
dispatch_context = autotvm.apply_history_best("conv2d.log")
best_config = dispatch_context.query(task.target, task.workload)
print("\nBest config:")
print(best_config)

# apply history best from log file
with autotvm.apply_history_best("conv2d.log"):
    with tvm.target.Target("cuda"):
        s, arg_bufs = conv2d_no_batching(N, H, W, CO, CI, KH, KW, strides, padding)
        func = tvm.build(s, arg_bufs)

# check correctness
a_np = np.random.uniform(size=(N, CI, H, W)).astype(np.float32)
w_np = np.random.uniform(size=(CO, CI, KH, KW)).astype(np.float32)
c_np = conv2d_nchw_python(a_np, w_np, strides, padding)

ctx = tvm.gpu()
a_tvm = tvm.nd.array(a_np, ctx=ctx)
w_tvm = tvm.nd.array(w_np, ctx=ctx)
c_tvm = tvm.nd.empty(c_np.shape, ctx=ctx)
func(a_tvm, w_tvm, c_tvm)

tvm.testing.assert_allclose(c_np, c_tvm.asnumpy(), rtol=1e-2)

# Evaluate running time. Here we choose a large repeat number (400) to reduce the noise
# and the overhead of kernel launch. You can also use nvprof to validate the result.
evaluator = func.time_evaluator(func.entry_name, ctx, number=400)
print("Time cost of this operator: %f" % evaluator(a_tvm, w_tvm, c_tvm).mean)

Out:

Best config:
[('tile_f', [-1, 2, 64, 1]), ('tile_y', [-1, 1, 1, 7]), ('tile_x', [-1, 1, 7, 1]), ('tile_rc', [-1, 2, 2]), ('tile_ry', [-1, 3, 1]), ('tile_rx', [-1, 1, 3]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],None,4881186
Time cost of this operator: 0.000568

Gallery generated by Sphinx-Gallery