28036549049

Committed 23 Jun 2026 03:20PM UTC coverage: 61.52% (-0.3%) from 61.779%

Build # 28036549049

Build Type

Pull #808

github

Committed by

web-flow

Commit Message

Merge 2133a817f into 6ee760cce

Pull Request Pull Request #808: enable batched gemm and memcpy on AMD GPUs analogous to the existing CUDA implementation

Coverage Stats

60 of 332 new or added lines in 7 files covered. (18.07%)

5 existing lines in 2 files now uncovered.

37177 of 60431 relevant lines covered (61.52%)

1008.85 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

72.33

/python/docc/compiler/docc_program.py

from abc import ABC, abstractmethod
import sys
from typing import Any, Dict, Optional
import json
import os
import re

from docc.sdfg import StructuredSDFG, TargetOptions
from docc.sdfg._sdfg import (
    _enable_statistics,
    _statistics_enabled_by_env,
    _statistics_summary,
)
from docc.compiler.compiled_sdfg import CompiledSDFG
from docc.compiler.target_registry import (
    get_target_schedule_fn,
    get_target_compile_fn,
    get_target_expand_fn,
    register_target_overrides,
)


def _parse_docc_debug() -> dict[str, str]:
    debug_env = os.environ.get("DOCC_DEBUG", "")
    debug_dict = {}
    if debug_env:
        for entry in re.split(r"[;:]", debug_env):
            if not entry:
                continue
            parts = entry.split("=", 1)
            key = parts[0].strip()
            value = parts[1].strip() if len(parts) > 1 else ""
            debug_dict[key] = value
    return debug_dict


def _is_debug_dump(flags: dict[str, str]) -> bool:
    return "dump" in flags


def _is_debug_compile(flags: dict[str, str]) -> bool:
    return "build" in flags


def _get_build_thread_count(flags: dict[str, str]) -> int:
    return int(flags.get("build_threads", "0"))


class DoccProgram(ABC):

    def __init__(
        self,
        name: str,
        target: str = "none",
        category: str = "server",
        instrumentation_mode: Optional[str] = None,
        capture_args: Optional[bool] = None,
        remote_tuning: bool = False,
    ):
        self.name = name
        self.target = target
        self.category = category
        self.remote_tuning = remote_tuning
        self.last_sdfg: Optional[StructuredSDFG] = None
        self._device_resident: bool = False
        self._device_backend: Optional[str] = None
        self.cache: dict = {}
        debug_flags = _parse_docc_debug()
        self.debug_dump: bool = _is_debug_dump(debug_flags)
        self.debug_build: bool = _is_debug_compile(debug_flags)
        self.build_thread_count: int = _get_build_thread_count(debug_flags)

        # Check environment variable DOCC_CI
        docc_ci = os.environ.get("DOCC_CI", "")
        if docc_ci:
            if docc_ci == "regions":
                if instrumentation_mode is None:
                    instrumentation_mode = "ols"
            elif docc_ci == "arg-capture":
                if capture_args is None:
                    capture_args = True
            else:
                # Full mode (or unknown value treated as full)
                if instrumentation_mode is None:
                    instrumentation_mode = "ols"
                if capture_args is None:
                    capture_args = True

        self.instrumentation_mode = instrumentation_mode
        self.capture_args = capture_args

    @abstractmethod
    def __call__(self, *args: Any) -> Any:
        pass

    @abstractmethod
    def compile(self, *args: Any, output_folder: Optional[str] = None) -> CompiledSDFG:
        pass

    def _resolve_compile_options(
        self,
        instrumentation_mode: Optional[str] = None,
        capture_args: Optional[bool] = None,
        remote_tuning: Optional[bool] = None,
    ) -> tuple[str, bool, bool]:
        """Resolve compile-time options, falling back to instance defaults and env vars."""
        if instrumentation_mode is None:
            instrumentation_mode = self.instrumentation_mode
        if capture_args is None:
            capture_args = self.capture_args
        if remote_tuning is None:
            remote_tuning = self.remote_tuning

        # Check environment variable DOCC_CI
        docc_ci = os.environ.get("DOCC_CI", "")
        if docc_ci:
            if docc_ci == "regions":
                if instrumentation_mode is None:
                    instrumentation_mode = "ols"
            elif docc_ci == "arg-capture":
                if capture_args is None:
                    capture_args = True
            else:
                # Full mode (or unknown value treated as full)
                if instrumentation_mode is None:
                    instrumentation_mode = "ols"
                if capture_args is None:
                    capture_args = True

        # Defaults
        if instrumentation_mode is None:
            instrumentation_mode = ""
        if capture_args is None:
            capture_args = False

        return instrumentation_mode, capture_args, remote_tuning

    def sdfg_pipe(
        self,
        sdfg: StructuredSDFG,
        output_folder: Optional[str],
        instrumentation_mode: str,
        capture_args: bool,
        remote_tuning: Optional[bool] = None,
    ) -> str:

        if self.debug_dump:
            sdfg.dump(output_folder, "py0.parsed", dump_dot=True)

        # Enable statistics if envvar is set
        if _statistics_enabled_by_env():
            _enable_statistics()

        sdfg.validate()

        if remote_tuning is None:
            remote_tuning = self.remote_tuning

        target_options = TargetOptions()
        target_options.target = self.target
        target_options.category = self.category
        target_options.remote_tuning = remote_tuning

        # Einsum detection
        sdfg.einsum()
        if self.debug_dump:
            sdfg.dump(output_folder, "py1.einsum", dump_dot=True)

        # Tensor targets keep tensor nodes
        custom_expand_fn = get_target_expand_fn(self.target)
        if custom_expand_fn is not None:
            custom_expand_fn(sdfg, self.category, {})
        else:
            sdfg.expand(target_options)
        if self.debug_dump:
            sdfg.dump(output_folder, "py2.expanded", dump_dot=True)

        # Simplify pipelines
        sdfg.simplify()
        if self.debug_dump:
            sdfg.dump(output_folder, "py3.opt", dump_dot=True)

        # Normalization for scheduling
        if self.target != "none":
            sdfg.normalize()

        if self.debug_dump or instrumentation_mode or capture_args:
            sdfg.dump(
                output_folder,
                "py4.norm",
                dump_dot=self.debug_dump,
                dump_json=True,
                record_for_instrumentation=True,
            )

        # Schedule if target is specified

        custom_schedule_fn = get_target_schedule_fn(self.target)
        if custom_schedule_fn is not None:
            custom_schedule_fn(sdfg, self.category, {"remote_tuning": remote_tuning})
        else:
            sdfg.schedule(target_options)

        if self.debug_dump:
            sdfg.dump(output_folder, "py5.post_sched", dump_dot=True)

        # Promote pointer arguments to device residency when the whole program keeps
        # data on device. Only enabled for the CUDA target: the device-resident
        # runtime path marshals arguments via cupy (CUDA), so promoting on other
        # backends (e.g. rocm) produces a device-resident ABI binary that the
        # runtime cannot feed correctly, causing GPU memory faults. Communicated
        # explicitly via the pass return value (bool), not through SDFG metadata.
        self._device_resident = False
        self._device_backend = None
        if self.target == "cuda":
            if sdfg.promote_device_residency(False):
                self._device_resident = True
                self._device_backend = self.target

        self.last_sdfg = sdfg

        custom_compile_fn = get_target_compile_fn(self.target)
        if custom_compile_fn is not None:
            lib_path = custom_compile_fn(
                sdfg,
                output_folder,
                instrumentation_mode,
                capture_args,
                {"debug_build": self.debug_build, "threads": self.build_thread_count},
            )
        else:
            lib_path = sdfg._compile(
                output_folder=output_folder,
                target=self.target,
                instrumentation_mode=instrumentation_mode,
                capture_args=capture_args,
                debug_build=self.debug_build,
                threads=self.build_thread_count,
            )

        # Dump statistics after compile
        if _statistics_enabled_by_env():
            print(_statistics_summary(), file=sys.stderr)

        # Record the device-residency decision in the persisted (py4.norm) SDFG
        # metadata. It is computed here (not stored in metadata by the pass) and
        # decides host vs device argument marshalling. Binary-reuse paths
        # (DOCC_REUSE_BINARIES) load only the cached .so + normalized SDFG and
        # never re-run scheduling/promotion, so without this they default to
        # host execution and feed host pointers into a device-resident binary
        # -> heap corruption / double free.
        if output_folder:
            self._persist_device_residency(output_folder, sdfg)

        return lib_path

    def _persist_device_residency(
        self, output_folder: str, sdfg: StructuredSDFG
    ) -> None:
        """Stamp the device-residency decision into the persisted SDFG metadata.

        Patches only the ``metadata`` object of the already-written
        ``py4.norm.json`` (the file the reuse path loads), leaving the SDFG
        structure and element IDs untouched so instrumentation references stay
        valid.
        """
        json_path = os.path.join(output_folder, f"{sdfg.name}.py4.norm.json")
        try:
            with open(json_path) as f:
                data = json.load(f)
            metadata = data.setdefault("metadata", {})
            metadata["device_resident"] = "1" if self._device_resident else "0"
            metadata["device_backend"] = self._device_backend or ""
            with open(json_path, "w") as f:
                json.dump(data, f)
        except (OSError, ValueError):
            pass

    @abstractmethod
    def to_sdfg(self, *args: Any) -> StructuredSDFG:
        pass

    @abstractmethod
    def _convert_inputs(self, args: tuple) -> tuple:
        pass

    @abstractmethod
    def _convert_outputs(self, result: Any, original_args: tuple) -> Any:
        pass

    def _get_cache_key(self, *args: Any) -> str:
        return ""

1	from abc import ABC, abstractmethod	4✔
2	import sys	4✔
3	from typing import Any, Dict, Optional	4✔
4	import json	4✔
5	import os	4✔
6	import re	4✔
7
8	from docc.sdfg import StructuredSDFG, TargetOptions	4✔
9	from docc.sdfg._sdfg import (	4✔
10	_enable_statistics,
11	_statistics_enabled_by_env,
12	_statistics_summary,
13	)
14	from docc.compiler.compiled_sdfg import CompiledSDFG	4✔
15	from docc.compiler.target_registry import (	4✔
16	get_target_schedule_fn,
17	get_target_compile_fn,
18	get_target_expand_fn,
19	register_target_overrides,
20	)
21
22
23	def _parse_docc_debug() -> dict[str, str]:	4✔
24	debug_env = os.environ.get("DOCC_DEBUG", "")	4✔
25	debug_dict = {}	4✔
26	if debug_env:	4✔
27	for entry in re.split(r"[;:]", debug_env):	×
28	if not entry:	×
29	continue	×
30	parts = entry.split("=", 1)	×
31	key = parts[0].strip()	×
32	value = parts[1].strip() if len(parts) > 1 else ""	×
33	debug_dict[key] = value	×
34	return debug_dict	4✔
35
36
37	def _is_debug_dump(flags: dict[str, str]) -> bool:	4✔
38	return "dump" in flags	4✔
39
40
41	def _is_debug_compile(flags: dict[str, str]) -> bool:	4✔
42	return "build" in flags	4✔
43
44
45	def _get_build_thread_count(flags: dict[str, str]) -> int:	4✔
46	return int(flags.get("build_threads", "0"))	4✔
47
48
49	class DoccProgram(ABC):	4✔
50
51	def __init__(	4✔
52	self,
53	name: str,
54	target: str = "none",
55	category: str = "server",
56	instrumentation_mode: Optional[str] = None,
57	capture_args: Optional[bool] = None,
58	remote_tuning: bool = False,
59	):
60	self.name = name	4✔
61	self.target = target	4✔
62	self.category = category	4✔
63	self.remote_tuning = remote_tuning	4✔
64	self.last_sdfg: Optional[StructuredSDFG] = None	4✔
65	self._device_resident: bool = False	4✔
66	self._device_backend: Optional[str] = None	4✔
67	self.cache: dict = {}	4✔
68	debug_flags = _parse_docc_debug()	4✔
69	self.debug_dump: bool = _is_debug_dump(debug_flags)	4✔
70	self.debug_build: bool = _is_debug_compile(debug_flags)	4✔
71	self.build_thread_count: int = _get_build_thread_count(debug_flags)	4✔
72
73	# Check environment variable DOCC_CI
74	docc_ci = os.environ.get("DOCC_CI", "")	4✔
75	if docc_ci:	4✔
76	if docc_ci == "regions":	×
77	if instrumentation_mode is None:	×
78	instrumentation_mode = "ols"	×
79	elif docc_ci == "arg-capture":	×
80	if capture_args is None:	×
81	capture_args = True	×
82	else:
83	# Full mode (or unknown value treated as full)
84	if instrumentation_mode is None:	×
85	instrumentation_mode = "ols"	×
86	if capture_args is None:	×
87	capture_args = True	×
88
89	self.instrumentation_mode = instrumentation_mode	4✔
90	self.capture_args = capture_args	4✔
91
92	@abstractmethod	4✔
93	def __call__(self, *args: Any) -> Any:	4✔
94	pass	×
95
96	@abstractmethod	4✔
97	def compile(self, *args: Any, output_folder: Optional[str] = None) -> CompiledSDFG:	4✔
98	pass	×
99
100	def _resolve_compile_options(	4✔
101	self,
102	instrumentation_mode: Optional[str] = None,
103	capture_args: Optional[bool] = None,
104	remote_tuning: Optional[bool] = None,
105	) -> tuple[str, bool, bool]:
106	"""Resolve compile-time options, falling back to instance defaults and env vars."""
107	if instrumentation_mode is None:	4✔
108	instrumentation_mode = self.instrumentation_mode	4✔
109	if capture_args is None:	4✔
110	capture_args = self.capture_args	4✔
111	if remote_tuning is None:	4✔
112	remote_tuning = self.remote_tuning	4✔
113
114	# Check environment variable DOCC_CI
115	docc_ci = os.environ.get("DOCC_CI", "")	4✔
116	if docc_ci:	4✔
117	if docc_ci == "regions":	×
118	if instrumentation_mode is None:	×
119	instrumentation_mode = "ols"	×
120	elif docc_ci == "arg-capture":	×
121	if capture_args is None:	×
122	capture_args = True	×
123	else:
124	# Full mode (or unknown value treated as full)
125	if instrumentation_mode is None:	×
126	instrumentation_mode = "ols"	×
127	if capture_args is None:	×
128	capture_args = True	×
129
130	# Defaults
131	if instrumentation_mode is None:	4✔
132	instrumentation_mode = ""	4✔
133	if capture_args is None:	4✔
134	capture_args = False	4✔
135
136	return instrumentation_mode, capture_args, remote_tuning	4✔
137
138	def sdfg_pipe(	4✔
139	self,
140	sdfg: StructuredSDFG,
141	output_folder: Optional[str],
142	instrumentation_mode: str,
143	capture_args: bool,
144	remote_tuning: Optional[bool] = None,
145	) -> str:
146
147	if self.debug_dump:	4✔
148	sdfg.dump(output_folder, "py0.parsed", dump_dot=True)	×
149
150	# Enable statistics if envvar is set
151	if _statistics_enabled_by_env():	4✔
152	_enable_statistics()	×
153
154	sdfg.validate()	4✔
155
156	if remote_tuning is None:	4✔
157	remote_tuning = self.remote_tuning	×
158
159	target_options = TargetOptions()	4✔
160	target_options.target = self.target	4✔
161	target_options.category = self.category	4✔
162	target_options.remote_tuning = remote_tuning	4✔
163
164	# Einsum detection
165	sdfg.einsum()	4✔
166	if self.debug_dump:	4✔
167	sdfg.dump(output_folder, "py1.einsum", dump_dot=True)	×
168
169	# Tensor targets keep tensor nodes
170	custom_expand_fn = get_target_expand_fn(self.target)	4✔
171	if custom_expand_fn is not None:	4✔
172	custom_expand_fn(sdfg, self.category, {})	4✔
173	else:
174	sdfg.expand(target_options)	4✔
175	if self.debug_dump:	4✔
176	sdfg.dump(output_folder, "py2.expanded", dump_dot=True)	×
177
178	# Simplify pipelines
179	sdfg.simplify()	4✔
180	if self.debug_dump:	4✔
181	sdfg.dump(output_folder, "py3.opt", dump_dot=True)	×
182
183	# Normalization for scheduling
184	if self.target != "none":	4✔
185	sdfg.normalize()	4✔
186
187	if self.debug_dump or instrumentation_mode or capture_args:	4✔
188	sdfg.dump(	4✔
189	output_folder,
190	"py4.norm",
191	dump_dot=self.debug_dump,
192	dump_json=True,
193	record_for_instrumentation=True,
194	)
195
196	# Schedule if target is specified
197
198	custom_schedule_fn = get_target_schedule_fn(self.target)	4✔
199	if custom_schedule_fn is not None:	4✔
200	custom_schedule_fn(sdfg, self.category, {"remote_tuning": remote_tuning})	4✔
201	else:
202	sdfg.schedule(target_options)	4✔
203
204	if self.debug_dump:	4✔
205	sdfg.dump(output_folder, "py5.post_sched", dump_dot=True)	×
206
207	# Promote pointer arguments to device residency when the whole program keeps
208	# data on device. Only enabled for the CUDA target: the device-resident
209	# runtime path marshals arguments via cupy (CUDA), so promoting on other
210	# backends (e.g. rocm) produces a device-resident ABI binary that the
211	# runtime cannot feed correctly, causing GPU memory faults. Communicated
212	# explicitly via the pass return value (bool), not through SDFG metadata.
213	self._device_resident = False	4✔
214	self._device_backend = None	4✔
215	if self.target == "cuda":	4✔
NEW 216	if sdfg.promote_device_residency(False):	×
217	self._device_resident = True	×
218	self._device_backend = self.target	×
219
220	self.last_sdfg = sdfg	4✔
221
222	custom_compile_fn = get_target_compile_fn(self.target)	4✔
223	if custom_compile_fn is not None:	4✔
224	lib_path = custom_compile_fn(	4✔
225	sdfg,
226	output_folder,
227	instrumentation_mode,
228	capture_args,
229	{"debug_build": self.debug_build, "threads": self.build_thread_count},
230	)
231	else:
232	lib_path = sdfg._compile(	4✔
233	output_folder=output_folder,
234	target=self.target,
235	instrumentation_mode=instrumentation_mode,
236	capture_args=capture_args,
237	debug_build=self.debug_build,
238	threads=self.build_thread_count,
239	)
240
241	# Dump statistics after compile
242	if _statistics_enabled_by_env():	4✔
243	print(_statistics_summary(), file=sys.stderr)	×
244
245	# Record the device-residency decision in the persisted (py4.norm) SDFG
246	# metadata. It is computed here (not stored in metadata by the pass) and
247	# decides host vs device argument marshalling. Binary-reuse paths
248	# (DOCC_REUSE_BINARIES) load only the cached .so + normalized SDFG and
249	# never re-run scheduling/promotion, so without this they default to
250	# host execution and feed host pointers into a device-resident binary
251	# -> heap corruption / double free.
252	if output_folder:	4✔
253	self._persist_device_residency(output_folder, sdfg)	4✔
254
255	return lib_path	4✔
256
257	def _persist_device_residency(	4✔
258	self, output_folder: str, sdfg: StructuredSDFG
259	) -> None:
260	"""Stamp the device-residency decision into the persisted SDFG metadata.
261
262	Patches only the ``metadata`` object of the already-written
263	``py4.norm.json`` (the file the reuse path loads), leaving the SDFG
264	structure and element IDs untouched so instrumentation references stay
265	valid.
266	"""
267	json_path = os.path.join(output_folder, f"{sdfg.name}.py4.norm.json")	4✔
268	try:	4✔
269	with open(json_path) as f:	4✔
270	data = json.load(f)	4✔
271	metadata = data.setdefault("metadata", {})	4✔
272	metadata["device_resident"] = "1" if self._device_resident else "0"	4✔
273	metadata["device_backend"] = self._device_backend or ""	4✔
274	with open(json_path, "w") as f:	4✔
275	json.dump(data, f)	4✔
276	except (OSError, ValueError):	4✔
277	pass	4✔
278
279	@abstractmethod	4✔
280	def to_sdfg(self, *args: Any) -> StructuredSDFG:	4✔
281	pass	×
282
283	@abstractmethod	4✔
284	def _convert_inputs(self, args: tuple) -> tuple:	4✔
285	pass	×
286
287	@abstractmethod	4✔
288	def _convert_outputs(self, result: Any, original_args: tuple) -> Any:	4✔
289	pass	×
290
291	def _get_cache_key(self, *args: Any) -> str:	4✔
292	return ""	×

daisytuner / docc / 28036549049

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous