def run_rearrange_operators_benchmarks(ctx=mx.cpu(), dtype='float32', profiler='native', int64_tensor='off', warmup=25, runs=100):
"""Runs benchmarks with the given context and precision (dtype) for all the
rearrange operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
int64_tensor: str, default 'off'
Input tensor size to use for tests (if on, dimensions >= 2**32)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Fetch all array rearrange operators
mx_rearrange_ops = get_all_rearrange_operators()
# Run benchmarks
mx_rearrange_op_results = run_op_benchmarks(mx_rearrange_ops, dtype, ctx, profiler, int64_tensor, warmup, runs)
return mx_rearrange_op_results
def run_sorting_searching_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
profiler='native',
warmup=25,
runs=100):
"""Runs benchmarks with the given context and precision (dtype)for all the sorting and searching
operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Fetch all Random Sampling Operators
mx_sort_search_ops = get_all_sorting_searching_operators()
# Run benchmarks
mx_sort_search_op_results = run_op_benchmarks(mx_sort_search_ops, dtype,
ctx, profiler, warmup, runs)
return mx_sort_search_op_results
def run_optimizer_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
profiler='native',
warmup=25,
runs=100):
"""Runs benchmarks with the given context and precision (dtype) for all the neural network
optimizer update operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Fetch all optimizer operators
mx_optimizer_ops = get_all_optimizer_operators()
# Run benchmarks
mx_optimizer_op_results = run_op_benchmarks(mx_optimizer_ops, dtype, ctx,
profiler, warmup, runs)
return mx_optimizer_op_results
def run_mx_binary_element_wise_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
warmup=25,
runs=100):
"""Runs benchmarks with the given context and precision (dtype)for all the binary
element_wise operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
warmup: int, default 10
Number of times to run for warmup
runs: int, default 50
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Fetch all Binary Element_wise Operators
mx_binary_element_wise_ops = get_all_elemen_wise_binary_operators()
# Run benchmarks
mx_binary_op_results = run_op_benchmarks(mx_binary_element_wise_ops, dtype,
ctx, warmup, runs)
return mx_binary_op_results
def run_activation_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
profiler='native',
warmup=25,
runs=100):
"""Runs benchmarks with the given context and precision (dtype)for all the activation
operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Module to use for tracking benchmark excecution time
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Fetch all NN Activation Operators
mx_activation_ops = get_all_nn_activation_operators()
# Run benchmarks
mx_activation_op_results = run_op_benchmarks(mx_activation_ops, dtype, ctx,
profiler, warmup, runs)
return mx_activation_op_results
def run_mx_unary_operators_benchmarks(ctx=mx.cpu(), dtype='float32', profiler='native', int64_tensor='off', warmup=25, runs=100):
"""Runs benchmarks with the given context, precision (dtype), and input data size (int64_tensor) for all the unary
operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
int64_tensor: str, default 'off'
Input tensor size to use for tests (if on, dimensions >= 2**32)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
standard_inputs = [{"args": [(1024, 1024)],
"num_outputs":1},
{"args": [(10000, 1)],
"num_outputs":1}]
int64_tensor_inputs = [{"args": [(2**32, 1)],
"num_outputs":1}]
if int64_tensor == 'on':
inputs = int64_tensor_inputs
else:
inputs = standard_inputs
# Run amp_multicast as it needs data as positional argument
amp_multicast_benchmark = run_performance_test([getattr(MX_OP_MODULE, "amp_multicast")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=inputs,
warmup=warmup,
runs=runs)
# Fetch all Unary Operators
mx_unary_broadcast_ops = get_all_unary_operators()
# Run benchmarks
mx_unary_op_results = run_op_benchmarks(mx_unary_broadcast_ops, dtype, ctx, profiler, int64_tensor, warmup, runs)
return merge_map_list(amp_multicast_benchmark + [mx_unary_op_results])
def run_linalg_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
profiler='native',
warmup=25,
runs=100):
"""Runs benchmarks with the given context and precision (dtype) for all the linear algebra
operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Individual tests for ops with specific requirements on input data
# linalg_potrf requires a positive definite matrix as input
linalg_potrf_benchmark = run_performance_test(getattr(
MX_OP_MODULE, "linalg_potrf"),
run_backward=False,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{
"A": [[1, 0], [0, 1]]
}, {
"A": [[2, -1, 0],
[-1, 2, -1],
[0, -1, 2]]
}],
warmup=warmup,
runs=runs)
# Fetch all Linear Algebra Operators
mx_linalg_ops = get_all_linalg_operators()
# Run benchmarks
mx_linalg_op_results = run_op_benchmarks(mx_linalg_ops, dtype, ctx,
profiler, warmup, runs)
return merge_map_list(linalg_potrf_benchmark + [mx_linalg_op_results])
def run_activation_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
profiler='native',
int64_tensor='off',
warmup=25,
runs=100):
"""Runs benchmarks with the given context, precision (dtype), and input data size (int64_tensor) for all the activation
operators (relu, sigmoid, softmax) in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Module to use for tracking benchmark excecution time
int64_tensor: str, default 'off'
Input tensor size to use for tests (if on, dimensions >= 2**32)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Fetch all NN Activation Operators
mx_activation_ops = get_all_nn_activation_operators()
# Run benchmarks
mx_activation_op_results = run_op_benchmarks(mx_activation_ops, dtype, ctx,
profiler, int64_tensor,
warmup, runs)
return mx_activation_op_results
def run_mx_misc_operators_benchmarks(ctx=mx.cpu(), dtype='float32', profiler='native', warmup=25, runs=100):
"""Runs benchmarks with the given context and precision (dtype) for all the miscellaneous
operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Individual tests for ops with positional args
array_ops_benchmark = run_performance_test([getattr(MX_OP_MODULE, "reset_arrays"),
getattr(MX_OP_MODULE, "multi_all_finite"),
getattr(MX_OP_MODULE, "multi_sum_sq")],
run_backward=False,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{"args": [(1024, 1024)],
"num_arrays": 1},
{"args": [(10000, 1)],
"num_arrays": 1},
{"args": [(10000, 10)],
"num_arrays": 1}],
warmup=warmup,
runs=runs)
add_n_benchmark = run_performance_test([getattr(MX_OP_MODULE, "add_n")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{"args": [(1024, 1024)]},
{"args": [(10000, 1)]},
{"args": [(10000, 10)]}],
warmup=warmup,
runs=runs)
# There are currently issus with UpSampling with bilinear interpolation.
# track issue here: https://github.com/apache/incubator-mxnet/issues/9138
upsampling_benchmark = run_performance_test([getattr(MX_OP_MODULE, "UpSampling")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{"args": (32, 3, 256, 256),
"scale": 2,
"sample_type": "nearest"},
{"args": (32, 3, 10000, 1),
"scale": 4,
"sample_type": "nearest"}],
warmup=warmup,
runs=runs)
# Create and register CustomAddOne operator for use in Custom op testing
c = CustomAddOneProp()
c.create_operator(ctx, [(1024,1024)], [dtype])
custom_benchmark = run_performance_test([getattr(MX_OP_MODULE, "Custom")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{"args": [(1024, 1024)],
"op_type": "CustomAddOne"},
{"args": [(10000, 1)],
"op_type": "CustomAddOne"},
{"args": [(10000, 10)],
"op_type": "CustomAddOne"}],
warmup=warmup,
runs=runs)
# Fetch remaining Miscellaneous Operators
mx_misc_ops = get_remaining_miscellaneous_operators()
# Run benchmarks
mx_misc_op_results = run_op_benchmarks(mx_misc_ops, dtype, ctx, profiler, warmup, runs)
return merge_map_list(array_ops_benchmark + add_n_benchmark + upsampling_benchmark + custom_benchmark + [mx_misc_op_results])
def run_optimizer_operators_benchmarks(ctx=mx.cpu(), dtype='float32', profiler='native', warmup=25, runs=100):
"""Runs benchmarks with the given context and precision (dtype) for all the neural network
optimizer update operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
# Run independent tests for ops that need specific input data
multi_mp_sgd_mom_res = run_performance_test([getattr(MX_OP_MODULE, "multi_mp_sgd_mom_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)),
"args1": nd.random_normal(shape=(5,5)), "args2": nd.random_normal(shape=(5,5)),
"args3": nd.random_normal(shape=(5,5)), "lrs": 0.1, "wds": 0.2,
"out": nd.random_normal(shape=(5,5))}],run_backward=False)
multi_sgd_mom_res = run_performance_test([getattr(MX_OP_MODULE, "multi_sgd_mom_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)),
"args1": nd.random_normal(shape=(5,5)),"args2": nd.random_normal(shape=(5,5)),
"lrs": 0.1, "wds": 0.2, "out": nd.random_normal(shape=(5,5))}], run_backward=False)
multi_sgd_res = run_performance_test([getattr(MX_OP_MODULE, "multi_sgd_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)),
"args1": nd.random_normal(shape=(5,5)), "lrs": 0.1, "wds": 0.2,
"out": nd.random_normal(shape=(5,5))}], run_backward=False)
multi_mp_sgd_res = run_performance_test([getattr(MX_OP_MODULE, "multi_mp_sgd_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)),
"args1": nd.random_normal(shape=(5,5)),"args2": nd.random_normal(shape=(5,5)),
"lrs": 0.1, "wds": 0.2, "out": nd.random_normal(shape=(5,5))}], run_backward=False)
preloaded_multi_mp_sgd_res = run_performance_test(
[getattr(MX_OP_MODULE, "preloaded_multi_mp_sgd_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)),
"args1": nd.random_normal(shape=(5,5)), "args2": nd.random_normal(shape=(5,5)),
"args3": nd.random_normal(shape=(1)), "args4": nd.random_normal(shape=(1)),
"out": nd.random_normal(shape=(5,5))}], run_backward=False)
preloaded_multi_sgd_mom_res = run_performance_test(
[getattr(MX_OP_MODULE, "preloaded_multi_sgd_mom_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)),
"args1": nd.random_normal(shape=(5,5)), "args2": nd.random_normal(shape=(5,5)),
"args3": nd.random_normal(shape=(1)), "args4": nd.random_normal(shape=(1)),
"out": nd.random_normal(shape=(5,5))}], run_backward=False)
preloaded_multi_sgd_res = run_performance_test(
[getattr(MX_OP_MODULE, "preloaded_multi_sgd_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)), "args1": nd.random_normal(shape=(5,5)),
"args4": nd.random_normal(shape=(1)), "args5": nd.random_normal(shape=(1)),
"out": nd.random_normal(shape=(5,5))}], run_backward=False)
preloaded_multi_mp_sgd_mom_res = run_performance_test(
[getattr(MX_OP_MODULE, "preloaded_multi_mp_sgd_mom_update")],
inputs=[{"args0": nd.random_normal(shape=(5,5)), "args1": nd.random_normal(shape=(5,5)),
"args2": nd.random_normal(shape=(5,5)), "args3": nd.random_normal(shape=(5,5)),
"args4": nd.random_normal(shape=(1)), "args5": nd.random_normal(shape=(1)),
"out": nd.random_normal(shape=(5,5))}], run_backward=False)
# Fetch remaining optimizer operators
mx_optimizer_ops = get_all_optimizer_operators()
# Run benchmarks
mx_optimizer_op_results = run_op_benchmarks(mx_optimizer_ops, dtype, ctx, profiler, warmup, runs)
return merge_map_list(multi_sgd_mom_res + multi_sgd_mom_res + multi_sgd_res + multi_mp_sgd_res + preloaded_multi_mp_sgd_res +\
preloaded_multi_sgd_mom_res + preloaded_multi_mp_sgd_res + preloaded_multi_mp_sgd_mom_res +\
[mx_optimizer_op_results])
def run_nn_basic_operators_benchmarks(ctx=mx.cpu(),
dtype='float32',
profiler='native',
int64_tensor='off',
warmup=25,
runs=100):
"""Runs benchmarks with the given context, precision (dtype), and data size (int64_tensor) for all the basic neural network
operators in MXNet.
Parameters
----------
ctx: mx.ctx
Context to run benchmarks
dtype: str, default 'float32'
Precision to use for benchmarks
profiler: str, default 'native'
Type of Profiler to use (native/python)
int64_tensor: str, default 'off'
Input tensor size to use for tests (if on, dimensions >= 2**32)
warmup: int, default 25
Number of times to run for warmup
runs: int, default 100
Number of runs to capture benchmark results
Returns
-------
Dictionary of results. Key -> Name of the operator, Value -> Benchmark results.
"""
standard_data_list = [(1024, 4, 4)]
int64_tensor_data_list = [(2**28, 4, 4)]
if int64_tensor == 'on':
data_list = int64_tensor_data_list
else:
data_list = standard_data_list
for data in data_list:
rnn_relu_benchmark = run_performance_test(
[getattr(MX_OP_MODULE, "RNN")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{
"data": data,
"parameters": (7, ),
"state": (1, 4, 1),
"mode": "rnn_relu",
"state_size": 1,
"num_layers": 1
}],
warmup=warmup,
runs=runs)
rnn_tanh_benchmark = run_performance_test(
[getattr(MX_OP_MODULE, "RNN")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{
"data": data,
"parameters": (7, ),
"state": (1, 4, 1),
"mode": "rnn_tanh",
"state_size": 1,
"num_layers": 1
}],
warmup=warmup,
runs=runs)
rnn_lstm_benchmark = run_performance_test(
[getattr(MX_OP_MODULE, "RNN")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{
"data": data,
"parameters": (28, ),
"state": (1, 4, 1),
"state_cell": (1, 4, 1),
"mode": "lstm",
"state_size": 1,
"num_layers": 1
}],
warmup=warmup,
runs=runs)
rnn_gru_benchmark = run_performance_test(
[getattr(MX_OP_MODULE, "RNN")],
run_backward=True,
dtype=dtype,
ctx=ctx,
profiler=profiler,
inputs=[{
"data": data,
"parameters": (21, ),
"state": (1, 4, 1),
"mode": "gru",
"state_size": 1,
"num_layers": 1
}],
warmup=warmup,
runs=runs)
# Fetch all NN Basic Operators
mx_nn_basic_ops = get_all_nn_basic_operators()
# Run benchmarks
mx_nn_basic_op_results = run_op_benchmarks(mx_nn_basic_ops, dtype, ctx,
profiler, int64_tensor, warmup,
runs)
return merge_map_list(rnn_relu_benchmark + rnn_tanh_benchmark +
rnn_lstm_benchmark + rnn_gru_benchmark +
[mx_nn_basic_op_results])
