kernel-tuner 1.4.0__tar.gz → 1.6.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/PKG-INFO +10 -5
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/README.md +1 -1
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/accuracy.py +38 -30
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/compiler.py +22 -19
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/hypertuner.py +44 -34
- kernel_tuner-1.6.0/kernel_tuner/backends/julia.py +684 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_helper.jl +82 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_helper.py +274 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_helper_ext_amdgpu.jl +7 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_helper_ext_cuda.jl +5 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_helper_ext_metal.jl +37 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_helper_ext_oneapi.jl +1 -0
- kernel_tuner-1.6.0/kernel_tuner/backends/julia_load_cuda.jl +20 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/nvcuda.py +31 -3
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/opencl.py +6 -13
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/pycuda.py +4 -14
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/core.py +187 -91
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/interface.py +118 -58
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/amd.py +14 -16
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/hip.py +4 -2
- kernel_tuner-1.6.0/kernel_tuner/observers/julia.py +162 -0
- kernel_tuner-1.6.0/kernel_tuner/observers/julia_nvml.jl +151 -0
- kernel_tuner-1.6.0/kernel_tuner/observers/julia_nvml.py +452 -0
- kernel_tuner-1.6.0/kernel_tuner/observers/metal.py +322 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/nvml.py +104 -30
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/observer.py +3 -2
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/tegra.py +4 -4
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/runners/parallel.py +2 -4
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/runners/sequential.py +2 -1
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/runners/simulation.py +1 -1
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/searchspace.py +4 -3
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/common.py +2 -1
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/util.py +226 -91
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/utils/directives.py +93 -93
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/pyproject.toml +17 -12
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/conftest.py +1 -1
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/context.py +29 -20
- kernel_tuner-1.6.0/test/strategies/test_dual_annealing.py +32 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_backend.py +7 -1
- kernel_tuner-1.6.0/test/test_cache_file.json.gz +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_core.py +2 -5
- kernel_tuner-1.6.0/test/test_julia_functions.py +219 -0
- kernel_tuner-1.6.0/test/test_julia_nvml_observer.py +188 -0
- kernel_tuner-1.6.0/test/test_metal_observer.py +48 -0
- kernel_tuner-1.6.0/test/test_nvml_continuous_observer.py +94 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_runners.py +11 -2
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_searchspace.py +27 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_time_budgets.py +11 -3
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_toml_file.py +2 -2
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_util_functions.py +84 -67
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/utils/nvcuda.py +1 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/utils/test_directives.py +3 -8
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/LICENSE +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/backend.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/cupy.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/hip/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/hip/hip.py +1 -1
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/backends/hip/util.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/energy/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/energy/energy.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/file_utils.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/hyper.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/integration.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/kernelbuilder.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/compiler.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/cupy.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/ncu.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/nvcuda.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/opencl.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/pmt.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/powersensor.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/pycuda.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/observers/register.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/runners/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/runners/runner.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/schema/T1/1.0.0/input-schema.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/schema/T4/1.0.0/metadata-schema.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/schema/T4/1.0.0/results-schema.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/basinhopping.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/bayes_opt.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/brute_force.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/diff_evo.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/dual_annealing.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/firefly_algorithm.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/gen_adaptive_tabu_greywolf.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/gen_hybrid_vndx.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/genetic_algorithm.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/greedy_ils.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/greedy_mls.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/hillclimbers.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/minimize.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/mls.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/ordered_greedy_mls.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/pso.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/pyatf_strategies.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/pymoo_minimize.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/random_sample.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/simulated_annealing.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/skopt.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/strategies/wrapper.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/utils/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/kernel_tuner/utils/nvcuda.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/convolution.cu +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_bayesian_optimization.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_cache_file.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_cache_time_energy.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_common.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_diff_evo.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_genetic_algorithm.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_pymoo_strategies.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/strategies/test_strategies.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/synthetic_fp32_cache_NVIDIA_RTX_A4000.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_T1_input.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_accuracy.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_cache_file.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_cache_file_T1_input.json +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_common.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_compiler_functions.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_cuda_functions.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_cupy_functions.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_custom_optimizer.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_energy.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_file_utils.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_hip_functions.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_hyper.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_integration.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_kernelbuilder.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_lazy_imports.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_nvml_mocked.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_observers.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_opencl_functions.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_pycuda_functions.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/test_pycuda_mocked.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/utils/__init__.py +0 -0
- {kernel_tuner-1.4.0 → kernel_tuner-1.6.0}/test/vector_add.cu +0 -0
|
@@ -1,13 +1,13 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: kernel_tuner
|
|
3
|
-
Version: 1.
|
|
3
|
+
Version: 1.6.0
|
|
4
4
|
Summary: An easy to use CUDA/OpenCL kernel tuner in Python
|
|
5
5
|
License-Expression: Apache-2.0
|
|
6
6
|
License-File: LICENSE
|
|
7
7
|
Keywords: auto-tuning,gpu,computing,pycuda,cuda,pyopencl,opencl
|
|
8
8
|
Author: Ben van Werkhoven
|
|
9
9
|
Author-email: b.vanwerkhoven@esciencecenter.nl
|
|
10
|
-
Requires-Python: >=3.
|
|
10
|
+
Requires-Python: >=3.11,<4
|
|
11
11
|
Classifier: Environment :: Console
|
|
12
12
|
Classifier: Environment :: GPU
|
|
13
13
|
Classifier: Development Status :: 5 - Production/Stable
|
|
@@ -21,19 +21,24 @@ Classifier: Operating System :: POSIX :: Linux
|
|
|
21
21
|
Classifier: Topic :: Scientific/Engineering
|
|
22
22
|
Classifier: Topic :: Software Development
|
|
23
23
|
Classifier: Topic :: System :: Distributed Computing
|
|
24
|
-
Classifier: Programming Language :: Python :: 3.10
|
|
25
24
|
Classifier: Programming Language :: Python :: 3.11
|
|
26
25
|
Classifier: Programming Language :: Python :: 3.12
|
|
27
26
|
Classifier: Programming Language :: Python :: 3.13
|
|
27
|
+
Classifier: Programming Language :: Python :: 3.14
|
|
28
|
+
Classifier: Programming Language :: Python :: 3.15
|
|
28
29
|
Provides-Extra: cuda
|
|
29
30
|
Provides-Extra: cuda-opencl
|
|
30
31
|
Provides-Extra: hip
|
|
32
|
+
Provides-Extra: julia
|
|
31
33
|
Provides-Extra: opencl
|
|
32
34
|
Provides-Extra: tutorial
|
|
33
35
|
Requires-Dist: cuda-python (>=12.6.0) ; extra == "cuda"
|
|
34
36
|
Requires-Dist: cuda-python (>=12.6.0) ; extra == "cuda-opencl"
|
|
35
37
|
Requires-Dist: hip-python ; extra == "hip"
|
|
36
38
|
Requires-Dist: jsonschema
|
|
39
|
+
Requires-Dist: juliacall (>=0.9.37) ; extra == "julia"
|
|
40
|
+
Requires-Dist: juliapkg (>=0.1.28) ; (python_version >= "3.14") and (extra == "julia")
|
|
41
|
+
Requires-Dist: juliapkg[zstd] (>=0.1.28) ; (python_version < "3.14") and (extra == "julia")
|
|
37
42
|
Requires-Dist: jupyter (>=1.0.0) ; extra == "tutorial"
|
|
38
43
|
Requires-Dist: matplotlib (>=3.5.0) ; extra == "tutorial"
|
|
39
44
|
Requires-Dist: numpy (>=2.0.0)
|
|
@@ -45,7 +50,7 @@ Requires-Dist: pymoo (>=0.6.1.6)
|
|
|
45
50
|
Requires-Dist: pynvml (>=11.4.1) ; extra == "cuda"
|
|
46
51
|
Requires-Dist: pyopencl ; extra == "cuda-opencl"
|
|
47
52
|
Requires-Dist: pyopencl ; extra == "opencl"
|
|
48
|
-
Requires-Dist: python-constraint2 (>=2.
|
|
53
|
+
Requires-Dist: python-constraint2 (>=2.5.0)
|
|
49
54
|
Requires-Dist: scikit-learn (>=1.0.2)
|
|
50
55
|
Requires-Dist: scipy (>=1.14.1)
|
|
51
56
|
Requires-Dist: xmltodict
|
|
@@ -74,7 +79,7 @@ Description-Content-Type: text/markdown
|
|
|
74
79
|
|
|
75
80
|
|
|
76
81
|
Create optimized GPU applications in any mainstream GPU
|
|
77
|
-
programming language (CUDA, HIP, OpenCL, OpenACC, OpenMP).
|
|
82
|
+
programming language (CUDA, HIP, OpenCL, OpenACC, OpenMP). For Julia, see [KernelTuner.jl](https://github.com/KernelTuner/KernelTuner.jl).
|
|
78
83
|
|
|
79
84
|
What Kernel Tuner does:
|
|
80
85
|
|
|
@@ -16,7 +16,7 @@
|
|
|
16
16
|
|
|
17
17
|
|
|
18
18
|
Create optimized GPU applications in any mainstream GPU
|
|
19
|
-
programming language (CUDA, HIP, OpenCL, OpenACC, OpenMP).
|
|
19
|
+
programming language (CUDA, HIP, OpenCL, OpenACC, OpenMP). For Julia, see [KernelTuner.jl](https://github.com/KernelTuner/KernelTuner.jl).
|
|
20
20
|
|
|
21
21
|
What Kernel Tuner does:
|
|
22
22
|
|
|
@@ -1,19 +1,25 @@
|
|
|
1
|
+
"""Module for accuracy measurement and tunable parameters."""
|
|
2
|
+
|
|
3
|
+
import logging
|
|
4
|
+
import re
|
|
1
5
|
from collections import UserDict
|
|
2
6
|
from typing import Dict
|
|
7
|
+
|
|
3
8
|
import numpy as np
|
|
4
|
-
import logging
|
|
5
|
-
import re
|
|
6
9
|
|
|
7
10
|
from kernel_tuner.observers import OutputObserver
|
|
8
11
|
|
|
9
12
|
|
|
10
13
|
class Tunable(UserDict):
|
|
14
|
+
"""``Tunable`` can be used as a parameter value dependent input argument when tuning kernels."""
|
|
15
|
+
|
|
11
16
|
def __init__(self, param_key: str, arrays: Dict):
|
|
12
|
-
"""The ``Tunable`` object can be used as an input argument when tuning
|
|
13
|
-
|
|
17
|
+
"""The ``Tunable`` object can be used as an input argument when tuning kernels.
|
|
18
|
+
|
|
19
|
+
It is a container that holds several arrays internally and
|
|
14
20
|
selects one array during benchmarking based on the value of a tunable parameter.
|
|
15
21
|
|
|
16
|
-
Example
|
|
22
|
+
Example:
|
|
17
23
|
-------
|
|
18
24
|
Consider this example::
|
|
19
25
|
|
|
@@ -37,6 +43,7 @@ class Tunable(UserDict):
|
|
|
37
43
|
self.param_key = param_key
|
|
38
44
|
|
|
39
45
|
def select_for_configuration(self, params):
|
|
46
|
+
"""Select the array for the given configuration."""
|
|
40
47
|
if callable(self.param_key):
|
|
41
48
|
option = self.param_key(params)
|
|
42
49
|
elif self.param_key in params:
|
|
@@ -45,14 +52,16 @@ class Tunable(UserDict):
|
|
|
45
52
|
option = eval(self.param_key, params, params)
|
|
46
53
|
|
|
47
54
|
if option not in self.data:
|
|
48
|
-
|
|
49
|
-
raise KeyError(
|
|
50
|
-
f"'{option}' is not a valid parameter value, should be one of: {list}"
|
|
51
|
-
)
|
|
55
|
+
raise KeyError(f"'{option}' is not a valid parameter value, should be one of: {', '.join(map(str, self.data.keys()))}")
|
|
52
56
|
|
|
53
|
-
|
|
57
|
+
# continue recursively until we find a non-Tunable
|
|
58
|
+
el = self.data[option]
|
|
59
|
+
if isinstance(el, Tunable):
|
|
60
|
+
return el.select_for_configuration(params)
|
|
61
|
+
else:
|
|
62
|
+
return el
|
|
54
63
|
|
|
55
|
-
def __call__(self, params):
|
|
64
|
+
def __call__(self, params): # noqa: D102
|
|
56
65
|
return self.select_for_configuration(params)
|
|
57
66
|
|
|
58
67
|
|
|
@@ -70,6 +79,7 @@ def _find_bfloat16_if_available():
|
|
|
70
79
|
if dtype is None:
|
|
71
80
|
try:
|
|
72
81
|
from ml_dtypes import bfloat16
|
|
82
|
+
|
|
73
83
|
dtype = bfloat16
|
|
74
84
|
except ImportError:
|
|
75
85
|
pass
|
|
@@ -78,6 +88,7 @@ def _find_bfloat16_if_available():
|
|
|
78
88
|
if dtype is None:
|
|
79
89
|
try:
|
|
80
90
|
from jax.numpy import bfloat16
|
|
91
|
+
|
|
81
92
|
dtype = bfloat16
|
|
82
93
|
except ImportError:
|
|
83
94
|
pass
|
|
@@ -86,6 +97,7 @@ def _find_bfloat16_if_available():
|
|
|
86
97
|
if dtype is None:
|
|
87
98
|
try:
|
|
88
99
|
from tensorflow import bfloat16
|
|
100
|
+
|
|
89
101
|
dtype = bfloat16.as_numpy_dtype
|
|
90
102
|
except ImportError:
|
|
91
103
|
pass
|
|
@@ -100,9 +112,9 @@ def _find_bfloat16_if_available():
|
|
|
100
112
|
|
|
101
113
|
|
|
102
114
|
def _to_float_dtype(x: str) -> np.dtype:
|
|
103
|
-
"""Convert a string to a numpy data type (``dtype``).
|
|
104
|
-
|
|
105
|
-
fallback.
|
|
115
|
+
"""Convert a string to a numpy data type (``dtype``).
|
|
116
|
+
|
|
117
|
+
This function recognizes common names (such as ``f16`` or ``kfloat``), and uses ``np.dtype(x)`` as a fallback.
|
|
106
118
|
"""
|
|
107
119
|
if isinstance(x, str):
|
|
108
120
|
x = x.lower()
|
|
@@ -123,16 +135,17 @@ def _to_float_dtype(x: str) -> np.dtype:
|
|
|
123
135
|
|
|
124
136
|
|
|
125
137
|
class TunablePrecision(Tunable):
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
):
|
|
129
|
-
"""The ``Tunable`` object can be used as an input argument when tuning
|
|
130
|
-
|
|
138
|
+
"""``TunablePrecision`` can be used as a precision-level dependent input argument when tuning kernels."""
|
|
139
|
+
|
|
140
|
+
def __init__(self, param_key: str, array: np.ndarray, dtypes: Dict[str, np.dtype] = None):
|
|
141
|
+
"""The ``Tunable`` object can be used as an input argument when tuning kernels.
|
|
142
|
+
|
|
143
|
+
It is a container that internally holds several arrays
|
|
131
144
|
containing the same data, but stored in using different levels of
|
|
132
145
|
precision. During benchamrking, one array is selected based on the value
|
|
133
146
|
of the tunable parameter called ``param_key``.
|
|
134
147
|
|
|
135
|
-
Example
|
|
148
|
+
Example:
|
|
136
149
|
-------
|
|
137
150
|
Consider this example::
|
|
138
151
|
|
|
@@ -156,7 +169,6 @@ class TunablePrecision(Tunable):
|
|
|
156
169
|
if bfloat16 is not None:
|
|
157
170
|
dtypes["bfloat16"] = bfloat16
|
|
158
171
|
|
|
159
|
-
|
|
160
172
|
# If dtype is a list, convert it to a dictionary
|
|
161
173
|
if isinstance(dtypes, (list, tuple)):
|
|
162
174
|
dtypes = dict((name, _to_float_dtype(name)) for name in dtypes)
|
|
@@ -197,7 +209,6 @@ def error_metric_from_name(user_key, EPS=1e-8):
|
|
|
197
209
|
The value of `EPS` is used for relative errors to prevent division by zero.
|
|
198
210
|
``
|
|
199
211
|
"""
|
|
200
|
-
|
|
201
212
|
# Prepocess the provided name:
|
|
202
213
|
# - convert to lowercase
|
|
203
214
|
# - remove the word "error"
|
|
@@ -278,17 +289,13 @@ def error_metric_from_name(user_key, EPS=1e-8):
|
|
|
278
289
|
raise ValueError(f"invalid error metric provided: {user_key}")
|
|
279
290
|
|
|
280
291
|
# cast both arguments to f64 before passing them to the metric
|
|
281
|
-
return lambda a, b: metric(
|
|
282
|
-
a.astype(np.float64, copy=False), b.astype(np.float64, copy=False)
|
|
283
|
-
)
|
|
292
|
+
return lambda a, b: metric(a.astype(np.float64, copy=False), b.astype(np.float64, copy=False))
|
|
284
293
|
|
|
285
294
|
|
|
286
295
|
class AccuracyObserver(OutputObserver):
|
|
287
|
-
"""``AccuracyObserver`` measures the error on the output produced by a kernel
|
|
288
|
-
by comparing the output against a reference output.
|
|
296
|
+
"""``AccuracyObserver`` measures the error on the output produced by a kernel by comparing to a reference output.
|
|
289
297
|
|
|
290
|
-
By default, it uses the root mean-squared error (RMSE) and uses the
|
|
291
|
-
metric name ``"error"``.
|
|
298
|
+
By default, it uses the root mean-squared error (RMSE) and uses the metric name ``"error"``.
|
|
292
299
|
"""
|
|
293
300
|
|
|
294
301
|
def __init__(self, metric=None, key="error", *, atol=1e-8):
|
|
@@ -303,7 +310,6 @@ class AccuracyObserver(OutputObserver):
|
|
|
303
310
|
:param atol: The tolerance used in relative metrics to prevent
|
|
304
311
|
division by zero. It is ignored by absolute error metrics.
|
|
305
312
|
"""
|
|
306
|
-
|
|
307
313
|
# Default metric is RMSE
|
|
308
314
|
if not metric:
|
|
309
315
|
metric = "rmse"
|
|
@@ -317,6 +323,7 @@ class AccuracyObserver(OutputObserver):
|
|
|
317
323
|
self.result = None
|
|
318
324
|
|
|
319
325
|
def process_output(self, answers, outputs):
|
|
326
|
+
"""Process the output produced by the kernel and compare it to the reference answers."""
|
|
320
327
|
errors = []
|
|
321
328
|
|
|
322
329
|
for answer, output in zip(answers, outputs):
|
|
@@ -326,4 +333,5 @@ class AccuracyObserver(OutputObserver):
|
|
|
326
333
|
self.result = max(errors)
|
|
327
334
|
|
|
328
335
|
def get_results(self):
|
|
336
|
+
"""Get the results produced by this observer."""
|
|
329
337
|
return dict([(self.key, self.result)])
|
|
@@ -1,13 +1,13 @@
|
|
|
1
|
-
"""
|
|
1
|
+
"""This module contains the functionality for running and compiling C functions"""
|
|
2
2
|
|
|
3
|
-
|
|
4
|
-
import
|
|
5
|
-
import platform
|
|
3
|
+
import _ctypes
|
|
4
|
+
import ctypes as C
|
|
6
5
|
import errno
|
|
7
|
-
import re
|
|
8
6
|
import logging
|
|
9
|
-
import
|
|
10
|
-
import
|
|
7
|
+
import platform
|
|
8
|
+
import re
|
|
9
|
+
import subprocess
|
|
10
|
+
from collections import namedtuple
|
|
11
11
|
|
|
12
12
|
import numpy as np
|
|
13
13
|
import numpy.ctypeslib
|
|
@@ -15,10 +15,9 @@ import numpy.ctypeslib
|
|
|
15
15
|
from kernel_tuner.backends.backend import CompilerBackend
|
|
16
16
|
from kernel_tuner.observers.compiler import CompilerRuntimeObserver
|
|
17
17
|
from kernel_tuner.util import (
|
|
18
|
-
get_temp_filename,
|
|
19
18
|
delete_temp_file,
|
|
19
|
+
get_temp_filename,
|
|
20
20
|
write_file,
|
|
21
|
-
SkippableFailure,
|
|
22
21
|
)
|
|
23
22
|
|
|
24
23
|
try:
|
|
@@ -88,7 +87,7 @@ class CompilerFunctions(CompilerBackend):
|
|
|
88
87
|
"""Class that groups the code for running and compiling C functions"""
|
|
89
88
|
|
|
90
89
|
def __init__(self, iterations=7, compiler_options=None, compiler=None, observers=None):
|
|
91
|
-
"""
|
|
90
|
+
"""Instantiate CFunctions object used for interacting with C code
|
|
92
91
|
|
|
93
92
|
:param iterations: Number of iterations used while benchmarking a kernel, 7 by default.
|
|
94
93
|
:type iterations: int
|
|
@@ -146,7 +145,7 @@ class CompilerFunctions(CompilerBackend):
|
|
|
146
145
|
self.name = platform.processor()
|
|
147
146
|
|
|
148
147
|
def ready_argument_list(self, arguments):
|
|
149
|
-
"""
|
|
148
|
+
"""Ready argument list to be passed to the C function
|
|
150
149
|
|
|
151
150
|
:param arguments: List of arguments to be passed to the C function.
|
|
152
151
|
The order should match the argument list on the C function.
|
|
@@ -181,7 +180,7 @@ class CompilerFunctions(CompilerBackend):
|
|
|
181
180
|
return ctype_args
|
|
182
181
|
|
|
183
182
|
def compile(self, kernel_instance):
|
|
184
|
-
"""
|
|
183
|
+
"""Call the C compiler to compile the kernel, return the function
|
|
185
184
|
|
|
186
185
|
:param kernel_instance: An object representing the specific instance of the tunable kernel
|
|
187
186
|
in the parameter space.
|
|
@@ -311,29 +310,33 @@ class CompilerFunctions(CompilerBackend):
|
|
|
311
310
|
def start_event(self):
|
|
312
311
|
"""Records the event that marks the start of a measurement
|
|
313
312
|
|
|
314
|
-
C backend does not use events
|
|
313
|
+
C backend does not use events
|
|
314
|
+
"""
|
|
315
315
|
pass
|
|
316
316
|
|
|
317
317
|
def stop_event(self):
|
|
318
318
|
"""Records the event that marks the end of a measurement
|
|
319
319
|
|
|
320
|
-
C backend does not use events
|
|
320
|
+
C backend does not use events
|
|
321
|
+
"""
|
|
321
322
|
pass
|
|
322
323
|
|
|
323
324
|
def kernel_finished(self):
|
|
324
325
|
"""Returns True if the kernel has finished, False otherwise
|
|
325
326
|
|
|
326
|
-
C backend does not support asynchronous launches
|
|
327
|
+
C backend does not support asynchronous launches
|
|
328
|
+
"""
|
|
327
329
|
return True
|
|
328
330
|
|
|
329
331
|
def synchronize(self):
|
|
330
332
|
"""Halts execution until device has finished its tasks
|
|
331
333
|
|
|
332
|
-
C backend does not support asynchronous launches
|
|
334
|
+
C backend does not support asynchronous launches
|
|
335
|
+
"""
|
|
333
336
|
pass
|
|
334
337
|
|
|
335
338
|
def run_kernel(self, func, c_args, threads, grid, stream=None):
|
|
336
|
-
"""
|
|
339
|
+
"""Runs the kernel once, returns whatever the kernel returns
|
|
337
340
|
|
|
338
341
|
:param func: A C function compiled for this specific configuration
|
|
339
342
|
:type func: ctypes._FuncPtr
|
|
@@ -367,7 +370,7 @@ class CompilerFunctions(CompilerBackend):
|
|
|
367
370
|
return time
|
|
368
371
|
|
|
369
372
|
def memset(self, allocation, value, size):
|
|
370
|
-
"""
|
|
373
|
+
"""Set the memory in allocation to the value in value
|
|
371
374
|
|
|
372
375
|
:param allocation: An Argument for some memory allocation unit
|
|
373
376
|
:type allocation: Argument
|
|
@@ -419,7 +422,7 @@ class CompilerFunctions(CompilerBackend):
|
|
|
419
422
|
self.memcpy_dtoh(arg, self.allocations[i])
|
|
420
423
|
|
|
421
424
|
def cleanup_lib(self):
|
|
422
|
-
"""
|
|
425
|
+
"""Unload the previously loaded shared library"""
|
|
423
426
|
if self.lib is None:
|
|
424
427
|
return
|
|
425
428
|
|
|
@@ -27,12 +27,14 @@ class ScoreObserver(BenchmarkObserver):
|
|
|
27
27
|
self.scores.append(self.dev.last_score)
|
|
28
28
|
|
|
29
29
|
def get_results(self):
|
|
30
|
-
results = {
|
|
30
|
+
results = {"score": mean(self.scores), "scores": self.scores.copy()}
|
|
31
31
|
self.scores = []
|
|
32
32
|
return results
|
|
33
33
|
|
|
34
|
+
|
|
34
35
|
class HypertunerFunctions(Backend):
|
|
35
36
|
"""Class for executing hyperparameter tuning."""
|
|
37
|
+
|
|
36
38
|
units = {}
|
|
37
39
|
|
|
38
40
|
def __init__(self, iterations, compiler_options=None):
|
|
@@ -51,30 +53,30 @@ class HypertunerFunctions(Backend):
|
|
|
51
53
|
"name": "dedispersion_milo",
|
|
52
54
|
"folder": folder,
|
|
53
55
|
"input_file": "dedispersion_milo.json",
|
|
54
|
-
"objective_performance_keys": ["time"]
|
|
56
|
+
"objective_performance_keys": ["time"],
|
|
55
57
|
},
|
|
56
58
|
{
|
|
57
59
|
"name": "hotspot_milo",
|
|
58
60
|
"folder": folder,
|
|
59
61
|
"input_file": "hotspot_milo.json",
|
|
60
|
-
"objective_performance_keys": ["GFLOP/s"]
|
|
62
|
+
"objective_performance_keys": ["GFLOP/s"],
|
|
61
63
|
},
|
|
62
64
|
{
|
|
63
65
|
"name": "convolution_milo",
|
|
64
66
|
"folder": folder,
|
|
65
67
|
"input_file": "convolution_milo.json",
|
|
66
|
-
"objective_performance_keys": ["time"]
|
|
68
|
+
"objective_performance_keys": ["time"],
|
|
67
69
|
},
|
|
68
70
|
{
|
|
69
71
|
"name": "gemm_milo",
|
|
70
72
|
"folder": folder,
|
|
71
73
|
"input_file": "gemm_milo.json",
|
|
72
|
-
"objective_performance_keys": ["time"]
|
|
73
|
-
}
|
|
74
|
+
"objective_performance_keys": ["time"],
|
|
75
|
+
},
|
|
74
76
|
]
|
|
75
77
|
# any additional settings
|
|
76
|
-
self.override = {
|
|
77
|
-
"experimental_groups_defaults": {
|
|
78
|
+
self.override = {
|
|
79
|
+
"experimental_groups_defaults": {
|
|
78
80
|
"repeats": 25,
|
|
79
81
|
"samples": self.iterations,
|
|
80
82
|
"minimum_fraction_of_budget_valid": 0.1,
|
|
@@ -84,10 +86,8 @@ class HypertunerFunctions(Backend):
|
|
|
84
86
|
"cutoff_percentile": 0.95,
|
|
85
87
|
"cutoff_percentile_start": 0.01,
|
|
86
88
|
"cutoff_type": "time",
|
|
87
|
-
"objective_time_keys": [
|
|
88
|
-
|
|
89
|
-
]
|
|
90
|
-
}
|
|
89
|
+
"objective_time_keys": ["all"],
|
|
90
|
+
},
|
|
91
91
|
}
|
|
92
92
|
|
|
93
93
|
# override the defaults with compiler options if provided
|
|
@@ -113,7 +113,7 @@ class HypertunerFunctions(Backend):
|
|
|
113
113
|
if arglist is None:
|
|
114
114
|
arglist = []
|
|
115
115
|
return arglist
|
|
116
|
-
|
|
116
|
+
|
|
117
117
|
def compile(self, kernel_instance):
|
|
118
118
|
super().compile(kernel_instance)
|
|
119
119
|
path = Path(__file__).parent.parent.parent / "hyperparamtuning"
|
|
@@ -121,37 +121,47 @@ class HypertunerFunctions(Backend):
|
|
|
121
121
|
|
|
122
122
|
# strategy settings
|
|
123
123
|
strategy: str = kernel_instance.arguments[0]
|
|
124
|
-
hyperparams = [{
|
|
124
|
+
hyperparams = [{"name": k, "value": v} for k, v in kernel_instance.params.items()]
|
|
125
125
|
hyperparams_string = "_".join(f"{k}={str(v)}" for k, v in kernel_instance.params.items())
|
|
126
|
-
searchspace_strategies = [
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
126
|
+
searchspace_strategies = [
|
|
127
|
+
{
|
|
128
|
+
"autotuner": "KernelTuner",
|
|
129
|
+
"name": f"{strategy.lower()}_{hyperparams_string}",
|
|
130
|
+
"display_name": strategy.replace("_", " ").capitalize(),
|
|
131
|
+
"search_method": strategy.lower(),
|
|
132
|
+
"search_method_hyperparameters": hyperparams,
|
|
133
|
+
}
|
|
134
|
+
]
|
|
133
135
|
|
|
134
136
|
name = kernel_instance.name if len(kernel_instance.name) > 0 else kernel_instance.kernel_source.kernel_name
|
|
135
|
-
experiments_filepath = generate_experiment_file(
|
|
136
|
-
|
|
137
|
+
experiments_filepath = generate_experiment_file(
|
|
138
|
+
name,
|
|
139
|
+
path,
|
|
140
|
+
searchspace_strategies,
|
|
141
|
+
self.applications,
|
|
142
|
+
self.gpus,
|
|
143
|
+
override=self.override,
|
|
144
|
+
generate_unique_file=True,
|
|
145
|
+
overwrite_existing_file=True,
|
|
146
|
+
)
|
|
137
147
|
return str(experiments_filepath)
|
|
138
|
-
|
|
148
|
+
|
|
139
149
|
def start_event(self):
|
|
140
150
|
return super().start_event()
|
|
141
|
-
|
|
151
|
+
|
|
142
152
|
def stop_event(self):
|
|
143
153
|
return super().stop_event()
|
|
144
|
-
|
|
154
|
+
|
|
145
155
|
def kernel_finished(self):
|
|
146
156
|
super().kernel_finished()
|
|
147
157
|
return True
|
|
148
|
-
|
|
158
|
+
|
|
149
159
|
def synchronize(self):
|
|
150
160
|
return super().synchronize()
|
|
151
|
-
|
|
161
|
+
|
|
152
162
|
def run_kernel(self, func, gpu_args=None, threads=None, grid=None, stream=None):
|
|
153
163
|
# from cProfile import Profile
|
|
154
|
-
|
|
164
|
+
|
|
155
165
|
# # generate the experiments file
|
|
156
166
|
# experiments_filepath = Path(func)
|
|
157
167
|
|
|
@@ -161,23 +171,23 @@ class HypertunerFunctions(Backend):
|
|
|
161
171
|
# pr.dump_stats('diff_evo_hypertune_hotspot.prof')
|
|
162
172
|
# self.last_score = scores[list(scores.keys())[0]]['score']
|
|
163
173
|
# raise ValueError(scores)
|
|
164
|
-
|
|
174
|
+
|
|
165
175
|
# generate the experiments file
|
|
166
176
|
experiments_filepath = Path(func)
|
|
167
177
|
|
|
168
178
|
# run the methodology to get a fitness score for this configuration
|
|
169
179
|
scores = get_strategy_scores(str(experiments_filepath), full_validate_on_load=False)
|
|
170
|
-
self.last_score = scores[list(scores.keys())[0]][
|
|
180
|
+
self.last_score = scores[list(scores.keys())[0]]["score"]
|
|
171
181
|
|
|
172
182
|
# remove the experiments file
|
|
173
183
|
experiments_filepath.unlink()
|
|
174
|
-
|
|
184
|
+
|
|
175
185
|
def memset(self, allocation, value, size):
|
|
176
186
|
return super().memset(allocation, value, size)
|
|
177
|
-
|
|
187
|
+
|
|
178
188
|
def memcpy_dtoh(self, dest, src):
|
|
179
189
|
return super().memcpy_dtoh(dest, src)
|
|
180
|
-
|
|
190
|
+
|
|
181
191
|
def memcpy_htod(self, dest, src):
|
|
182
192
|
return super().memcpy_htod(dest, src)
|
|
183
193
|
|