GPU heuristics determinism integration #10
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,249 @@ | ||
| <?xml version="1.0" encoding="utf-8"?> | ||
| <ComputeSanitizerOutput> | ||
| <record> | ||
| <kind>Initcheck</kind> | ||
| <what> | ||
| <text>Uninitialized __global__ memory read of size 4 bytes</text> | ||
| <size>4</size> | ||
| </what> | ||
| <where> | ||
| <func>.*</func> | ||
| </where> | ||
| <hostStack> | ||
| <frame> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>cusparseCsr2cscEx2</func> | ||
| <module>.*libcusparse.so.*</module> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <record> | ||
| <kind>Initcheck</kind> | ||
| <what> | ||
| <text>Uninitialized __global__ memory read of size 4 bytes</text> | ||
| <size>4</size> | ||
| </what> | ||
| <where> | ||
| <func>ThreadLoad</func> | ||
| </where> | ||
| <hostStack> | ||
| <frame> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>libcudart.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>cudaLaunchKernel</func> | ||
| </frame> | ||
| <frame> | ||
| <func>.*cub::.*::Device(Segmented)?(Reduce|Scan)(SingleTile)?Kernel.*</func> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <record> | ||
| <kind>Initcheck</kind> | ||
| <what> | ||
| <text>Uninitialized __global__ memory read of size 2 bytes</text> | ||
| <size>2</size> | ||
| </what> | ||
| <where> | ||
| <func>ThreadLoad</func> | ||
| </where> | ||
| <hostStack> | ||
| <frame> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>libcudart.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>cudaLaunchKernel</func> | ||
| </frame> | ||
| <frame> | ||
| <func>.*cub::.*::Device(Segmented)?(Reduce|Scan)(SingleTile)?Kernel.*</func> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <record> | ||
| <kind>Initcheck</kind> | ||
| <what> | ||
| <text>Uninitialized __global__ memory read of size 8 bytes</text> | ||
| <size>8</size> | ||
| </what> | ||
| <where> | ||
| <func>DeviceSegmentedReduceKernel</func> | ||
| </where> | ||
| </record> | ||
|
Comment on lines
+70
to
+79
There was a problem hiding this comment. 🧩 Analysis chain🏁 Script executed: #!/bin/bash
python - <<'PY'
import xml.etree.ElementTree as ET
p = "ci/compute-sanitizer-suppressions.xml"
root = ET.parse(p).getroot()
for i, rec in enumerate(root.findall("record"), 1):
kind = (rec.findtext("kind") or "").strip()
where_func = (rec.findtext("./where/func") or "").strip()
has_host_stack = rec.find("hostStack") is not None
if kind == "Initcheck" and (not has_host_stack):
print(f"record {i}: Initcheck without hostStack, where/func={where_func or '<none>'}")
PYRepository: aliceb-nv/cuopt-fork Length of output: 141 Add a The suppression for 🤖 Prompt for AI Agents |
||
| <record> | ||
| <kind>Initcheck</kind> | ||
| <what> | ||
| <text>Uninitialized __global__ memory read of size 4 bytes</text> | ||
| <size>4</size> | ||
| </what> | ||
| <where> | ||
| <func>ThreadLoad</func> | ||
| </where> | ||
| <hostStack> | ||
| <frame> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>libcudart.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>libcudart.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcuopt.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>.*Device(Reduce|Scan)Kernel.*</func> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <!-- Rule matching cccl's pattern of copying tuples back to host after reduce_by_keys, which contain uninitialized padding --> | ||
| <!-- Because of aggressive inlining, thrust calls are elided out of the host stack, which prevents a more finely grained rule. In practice this is good enough --> | ||
| <record> | ||
| <kind>InitcheckApiError</kind> | ||
| <level>Error</level> | ||
| <what> | ||
| <text>Host API uninitialized memory access</text> | ||
| <accessSize>16</accessSize> | ||
| </what> | ||
| <hostStack> | ||
| <frame> | ||
| <func>cuMemcpyDtoHAsync.*</func> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <!-- Suppress uninit copies on rmm::device_vector copy constructor - often vector members are allocated but not filled --> | ||
| <record> | ||
| <kind>InitcheckApiError</kind> | ||
| <level>Error</level> | ||
| <what> | ||
| <text>Host API uninitialized memory access</text> | ||
| </what> | ||
| <hostStack> | ||
| <frame> | ||
| <func>cuMemcpyAsync</func> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*librmm.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>rmm::device_buffer::device_buffer</func> | ||
| <module>.*librmm.so.*</module> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <!-- Suppress likely harmless Thrust/CUB tuple-buffer initcheck reads during sort_by_key | ||
| in trivial_presolve's COO->CSC reorder. Source and destination arrays are validated | ||
| immediately before the sort; the warning appears to come from internal tuple staging. --> | ||
| <record> | ||
| <kind>Initcheck</kind> | ||
| <what> | ||
| <text>Uninitialized __global__ memory read</text> | ||
| </what> | ||
| <where> | ||
| <func>transform_kernel</func> | ||
| </where> | ||
| <hostStack> | ||
| <frame> | ||
| <func>cuLaunchKernel_ptsz</func> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>cudaLaunchKernel_ptsz</func> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <record> | ||
| <kind>InitcheckApiError</kind> | ||
| <level>Error</level> | ||
| <what> | ||
| <text>Host API uninitialized memory access</text> | ||
| </what> | ||
| <hostStack> | ||
| <frame> | ||
| <func>cuMemcpyAsync</func> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*librmm.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*librmm.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>rmm::device_uvector.*::device_uvector</func> | ||
| <module>.*libcuopt.so.*</module> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| <!-- Uninitialized device-to-device copies are usually harmless - if actualy bogus, errors may be caught later on --> | ||
| <record> | ||
| <kind>InitcheckApiError</kind> | ||
| <level>Error</level> | ||
| <what> | ||
| <text>Host API uninitialized memory access</text> | ||
| </what> | ||
| <hostStack> | ||
| <frame> | ||
| <func>cuMemcpyDtoDAsync.*</func> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
|
Comment on lines
+210
to
+222
There was a problem hiding this comment. 🧩 Analysis chain🏁 Script executed: #!/bin/bash
python - <<'PY'
import xml.etree.ElementTree as ET, re
p = "ci/compute-sanitizer-suppressions.xml"
root = ET.parse(p).getroot()
for i, rec in enumerate(root.findall("record"), 1):
if (rec.findtext("kind") or "").strip() != "InitcheckApiError":
continue
frames = rec.findall("./hostStack/frame")
funcs = [(f.findtext("func") or "").strip() for f in frames]
if any(re.search(r"cuMemcpyDtoDAsync", fn) for fn in funcs):
print(f"record {i}: cuMemcpyDtoDAsync suppression has {len(frames)} host frame(s)")
for n, fn in enumerate(funcs, 1):
print(f" frame {n}: {fn or '<module-only frame>'}")
PYRepository: aliceb-nv/cuopt-fork Length of output: 155 🏁 Script executed: # First, let's look at the full context of the suppression file
head -250 ci/compute-sanitizer-suppressions.xml | tail -50Repository: aliceb-nv/cuopt-fork Length of output: 1363 🏁 Script executed: # Check if there are any comments in the XML that explain this suppression
grep -B5 -A5 "cuMemcpyDtoDAsync" ci/compute-sanitizer-suppressions.xmlRepository: aliceb-nv/cuopt-fork Length of output: 308 🏁 Script executed: # Look at the git history for when this suppression was added and why
git log --oneline -n 20 -- ci/compute-sanitizer-suppressions.xmlRepository: aliceb-nv/cuopt-fork Length of output: 84 🏁 Script executed: # Check if this suppression is mentioned in any documentation or comments in the codebase
rg "cuMemcpyDtoDAsync" --type txt --type md -A3 -B3Repository: aliceb-nv/cuopt-fork Length of output: 46 Constrain This rule broadly suppresses all 🤖 Prompt for AI Agents |
||
| <record> | ||
| <kind>InitcheckApiError</kind> | ||
| <level>Error</level> | ||
| <what> | ||
| <text>Host API uninitialized memory access</text> | ||
| </what> | ||
| <hostStack> | ||
| <frame> | ||
| <func>cuMemcpyAsync</func> | ||
| <module>.*libcuda.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <module>.*libcudart.so.*</module> | ||
| </frame> | ||
| <frame> | ||
| <func>cudaMemcpyAsync</func> | ||
| </frame> | ||
| <frame> | ||
| <func>rmm::device_buffer::resize</func> | ||
| <module>.*librmm.so.*</module> | ||
| </frame> | ||
| </hostStack> | ||
| </record> | ||
| </ComputeSanitizerOutput> | ||
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
🧩 Analysis chain
🏁 Script executed:
Repository: aliceb-nv/cuopt-fork
Length of output: 113
🏁 Script executed:
Repository: aliceb-nv/cuopt-fork
Length of output: 1762
🏁 Script executed:
Repository: aliceb-nv/cuopt-fork
Length of output: 433
🏁 Script executed:
# Get line count to understand file size wc -l benchmarks/linear_programming/cuopt/run_mip.cppRepository: aliceb-nv/cuopt-fork
Length of output: 115
🏁 Script executed:
Repository: aliceb-nv/cuopt-fork
Length of output: 1964
🏁 Script executed:
Repository: aliceb-nv/cuopt-fork
Length of output: 1228
Use
steady_clockfor elapsed-time capture.high_resolution_clockis not guaranteed to be monotonic and may aliassystem_clockon some implementations, causing the incumbent trace timestamps to jump backwards if system time is adjusted.steady_clockis guaranteed monotonic and is the correct choice for benchmark elapsed-time measurement. Apply this change to the constructor parameter, member variable, callback implementation, and thestart_run_solverassignment and finalendcalculation.Proposed changes
🤖 Prompt for AI Agents