Make tutorials build on a DGX Spark w/ CUDA 13.0 by gnurizen · Pull Request #3 · eunomia-bpf/cupti-tutorial

gnurizen · 2025-11-20T17:07:00Z

Untested on 12.* and not sure makefile includes will fly on windows.

Copilot

Pull request overview

Updates CUPTI sample tutorials to build with CUDA 13.0-era toolchains (e.g., DGX Spark) by adapting to the CUDA Driver API cuCtxCreate signature change and centralizing SM architecture flags in a shared Makefile include.

Changes:

Replace direct cuCtxCreate(...) calls with a COMPAT_cuCtxCreate(...) wrapper for CUDA 13 compatibility.
Consolidate SMS/-gencode handling across many samples via common/common.mk.
Update activity_trace to use $(CUDA_INSTALL_PATH) for nvcc rather than a hardcoded CUDA 12.9 path.

Reviewed changes

Copilot reviewed 50 out of 50 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
userrange_profiling/user_range_profiling.cu	Switch context creation to `COMPAT_cuCtxCreate`.
userrange_profiling/simplecuda.cu	Switch context creation to `COMPAT_cuCtxCreate`.
userrange_profiling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
unified_memory/Makefile	Replace per-sample SM list with shared `common/common.mk`.
sass_source_map/Makefile	Replace per-sample SM list with shared `common/common.mk`.
sass_metrics/sass_metrics.cu	Switch context creation to `COMPAT_cuCtxCreate`.
sass_metrics/Makefile	Replace per-sample SM list with shared `common/common.mk`.
range_profiling/range_profiling.cu	Switch context creation to `COMPAT_cuCtxCreate`.
range_profiling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
profiling_injection/complex_target.cu	Switch context creation to `COMPAT_cuCtxCreate`.
profiling_injection/Makefile	Replace per-sample SM list with shared `common/common.mk`.
pm_sampling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
pc_sampling_start_stop/pc_sampling_start_stop.cu	Switch context creation to `COMPAT_cuCtxCreate`.
pc_sampling_start_stop/Makefile	Replace per-sample SM list with shared `common/common.mk`.
pc_sampling/Makefile	Use `DEFAULT_SM_ARCH` and shared `common/common.mk` for arch selection.
nvlink_bandwidth/nvlink_bandwidth.cu	Switch context creation to `COMPAT_cuCtxCreate`.
nvlink_bandwidth/Makefile	Replace per-sample SM list with shared `common/common.mk`.
nested_range_profiling/nested_range_profiling.cu	Switch context creation to `COMPAT_cuCtxCreate`.
nested_range_profiling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
event_sampling/event_sampling.cu	Switch context creation to `COMPAT_cuCtxCreate`.
event_sampling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
event_multi_gpu/event_multi_gpu.cu	Switch context creation to `COMPAT_cuCtxCreate`.
cupti_nvtx_ext_payload/Makefile	Replace per-sample SM list with shared `common/common.mk`.
cupti_nvtx/cupti_nvtx.cu	Switch context creation to `COMPAT_cuCtxCreate`.
cupti_nvtx/Makefile	Replace per-sample SM list with shared `common/common.mk`.
cupti_metric_properties/cupti_metric_properties.cpp	Switch context creation to `COMPAT_cuCtxCreate`.
cupti_metric_properties/Makefile	Replace per-sample SM list with shared `common/common.mk`.
cupti_external_correlation/cupti_external_correlation.cu	Switch context creation to `COMPAT_cuCtxCreate`.
cupti_external_correlation/Makefile	Replace per-sample SM list with shared `common/common.mk`.
cupti_correlation/Makefile	Replace per-sample SM list with shared `common/common.mk`.
cuda_memory_trace/Makefile	Replace per-sample SM list with shared `common/common.mk`.
cuda_graphs_trace/Makefile	Replace per-sample SM list with shared `common/common.mk`.
concurrent_profiling/concurrent_profiling.cu	Switch context creation to `COMPAT_cuCtxCreate`.
concurrent_profiling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
common/helper_cupti.h	Add `COMPAT_cuCtxCreate` macro keyed off CUDA toolkit version.
common/common.mk	New shared Makefile logic for CUDA-version-based SM selection and `-gencode` generation.
checkpoint_kernels/checkpoint_kernels.cu	Switch context creation to `COMPAT_cuCtxCreate`.
checkpoint_kernels/Makefile	Replace per-sample SM list with shared `common/common.mk`.
callback_timestamp/callback_timestamp.cu	Switch context creation to `COMPAT_cuCtxCreate`.
callback_timestamp/Makefile	Replace per-sample SM list with shared `common/common.mk`.
callback_profiling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
callback_metric/callback_metric.cu	Switch context creation to `COMPAT_cuCtxCreate`.
callback_metric/Makefile	Replace per-sample SM list with shared `common/common.mk`.
callback_event/callback_event.cu	Switch context creation to `COMPAT_cuCtxCreate`.
callback_event/Makefile	Replace per-sample SM list with shared `common/common.mk`.
autorange_profiling/simplecuda.cu	Switch context creation to `COMPAT_cuCtxCreate`.
autorange_profiling/auto_range_profiling.cu	Switch context creation to `COMPAT_cuCtxCreate`.
autorange_profiling/Makefile	Replace per-sample SM list with shared `common/common.mk`.
activity_trace_async/Makefile	Replace per-sample SM list with shared `common/common.mk`.
activity_trace/Makefile	Switch `NVCC` from hardcoded CUDA 12.9 path to `$(CUDA_INSTALL_PATH)`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-02-11T01:08:11Z

+        DEFAULT_SM_ARCH := sm_52
+    endif
+endif
+


DEFAULT_SM_ARCH is only set inside the ifndef SMS block. If a user overrides SMS (common when targeting a single GPU), DEFAULT_SM_ARCH remains unset and samples like pc_sampling that use -arch=$(DEFAULT_SM_ARCH) will fail. Define DEFAULT_SM_ARCH independently when it is not provided (e.g., derive it from $(firstword $(SMS))).

Suggested change

# If SMS was provided by the user, ensure DEFAULT_SM_ARCH is also set.

# Derive it from the first entry in SMS when not explicitly specified.

ifndef DEFAULT_SM_ARCH

ifdef SMS

DEFAULT_SM_ARCH := sm_$(firstword $(SMS))

endif

endif

Copilot · 2026-02-11T01:08:11Z

+# Detect CUDA version to determine supported SM architectures
+# CUDA 13+ removed support for compute capabilities below sm_75
+CUDA_VERSION := $(shell $(CUDA_INSTALL_PATH)/bin/nvcc --version 2>/dev/null | grep release | sed 's/.*release //' | sed 's/,.*//' | cut -d. -f1)
+
+# Set default SMS based on CUDA version if not already set
+ifndef SMS
+    ifeq ($(shell test $(CUDA_VERSION) -ge 13 2>/dev/null; echo $$?),0)
+        # CUDA 13+ only supports sm_75 and above


CUDA version detection relies on Unix utilities (grep, sed, cut, test) in a pipeline. These are unlikely to work in the Windows_NT codepaths where GNU make may be using cmd.exe, which would break any sample Makefile that includes this file. Consider guarding auto-detection behind a non-Windows check and/or providing a simple fallback (require SMS to be set on Windows, or use a more portable parsing approach).

Copilot · 2026-02-11T01:08:11Z

+
+# Detect CUDA version to determine supported SM architectures
+# CUDA 13+ removed support for compute capabilities below sm_75
+CUDA_VERSION := $(shell $(CUDA_INSTALL_PATH)/bin/nvcc --version 2>/dev/null | grep release | sed 's/.*release //' | sed 's/,.*//' | cut -d. -f1)


CUDA_VERSION is derived from $(CUDA_INSTALL_PATH)/bin/nvcc directly. Several sample Makefiles allow overriding NVCC (or quote it), and CUDA_INSTALL_PATH might not match the actual nvcc used. Consider basing detection on $(NVCC) instead (or falling back to nvcc on PATH) so the architecture selection matches the compiler actually invoked.

Suggested change

CUDA_VERSION := $(shell $(CUDA_INSTALL_PATH)/bin/nvcc --version 2>/dev/null | grep release | sed 's/.*release //' | sed 's/,.*//' | cut -d. -f1)

NVCC ?= nvcc

CUDA_VERSION := $(shell $(NVCC) --version 2>/dev/null | grep release | sed 's/.*release //' | sed 's/,.*//' | cut -d. -f1)

Copilot · 2026-02-11T01:08:11Z

@@ -2,7 +2,7 @@
 # Copyright 2011-2013 NVIDIA Corporation. All rights reserved
 # 
 INCLUDES=-I../common


NVCC is now derived from $(CUDA_INSTALL_PATH), but this Makefile never defines CUDA_INSTALL_PATH (unlike most other samples). On a clean environment this will resolve to /bin/nvcc and break the build. Add a CUDA_INSTALL_PATH ?= ... default (or compute NVCC ?= nvcc from PATH) to keep the sample self-contained.

Suggested change

INCLUDES=-I../common

INCLUDES=-I../common

CUDA_INSTALL_PATH ?= /usr/local/cuda

yunwei37

The CUDA 13 compatibility direction is useful, but the current head still has four build-path blockers that need to be resolved before merge:

DEFAULT_SM_ARCH is undefined when callers override SMS.
CUDA-version detection uses Unix-only shell tools in the Windows_NT paths this PR still supports.
Version detection should follow the actual overridable NVCC, rather than hard-coding $(CUDA_INSTALL_PATH)/bin/nvcc.
activity_trace/Makefile now derives NVCC from CUDA_INSTALL_PATH without defining a default, which resolves to /bin/nvcc in a clean environment.

Please address the existing inline threads and add at least one CUDA 13 build result. A CUDA 12.x build result is also needed because the PR changes the shared Makefile included by many samples.

Make tutorials build on a DGX Spark w/ CUDA 13.0

ebc9e51

Untested on 12.* and not sure makefile includes will fly on windows.

pull-request-size Bot added the size/L label Nov 20, 2025

This was referenced Dec 1, 2025

Weekly Org Report (2025-11-17..2025-11-23) eunomia-bpf/eunomia.dev#56

Open

Monthly Org Report (2025-11-01..2025-11-30) eunomia-bpf/eunomia.dev#57

Open

yunwei37 requested a review from Copilot February 11, 2026 01:00

Copilot started reviewing on behalf of yunwei37 February 11, 2026 01:01 View session

Copilot AI reviewed Feb 11, 2026

View reviewed changes

yunwei37 requested changes Jul 19, 2026

View reviewed changes

yunwei37 mentioned this pull request Jul 19, 2026

Support for CUDA-13.0 #4

Open

Navigation Menu

Uh oh!

Uh oh!

Make tutorials build on a DGX Spark w/ CUDA 13.0 - #3

Make tutorials build on a DGX Spark w/ CUDA 13.0#3
gnurizen wants to merge 1 commit into
eunomia-bpf:masterfrom
gnurizen:cuda_13

gnurizen commented Nov 20, 2025 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Feb 11, 2026

Uh oh!

Copilot AI Feb 11, 2026

Uh oh!

Copilot AI Feb 11, 2026

Uh oh!

Copilot AI Feb 11, 2026

Uh oh!

yunwei37 left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

+# If SMS was provided by the user, ensure DEFAULT_SM_ARCH is also set.
+# Derive it from the first entry in SMS when not explicitly specified.
+ifndef DEFAULT_SM_ARCH
+    ifdef SMS
+        DEFAULT_SM_ARCH := sm_$(firstword $(SMS))
+    endif
+endif

	CUDA_VERSION := $(shell $(CUDA_INSTALL_PATH)/bin/nvcc --version 2>/dev/null \| grep release \| sed 's/.release //' \| sed 's/,.//' \| cut -d. -f1)
	NVCC ?= nvcc
	CUDA_VERSION := $(shell $(NVCC) --version 2>/dev/null \| grep release \| sed 's/.release //' \| sed 's/,.//' \| cut -d. -f1)

	INCLUDES=-I../common
	INCLUDES=-I../common
	CUDA_INSTALL_PATH ?= /usr/local/cuda

Uh oh!

Uh oh!

Conversation

gnurizen commented Nov 20, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Feb 11, 2026

Choose a reason for hiding this comment

Uh oh!

yunwei37 left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

gnurizen commented Nov 20, 2025 •

edited

Loading