NASP/cuda-samples
84
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-13 00:08:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

Release v13.3 of the CUDA samples with CUDA 13.3 Toolkit (#435)

Browse Source 
This is the release of the CUDA 13.3 samples, which include additions for CUDA Tile C++, and updated CCCL and Python samples.
This commit is contained in:
Dheemanth
2026-05-27 14:50:59 -07:00
committed by GitHub
parent aeab82ff30
commit b7c5481c55
150 changed files with 7977 additions and 246 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/2_CoreConcepts/streamingCopyComputeOverlap/README.md
View File

@@ -67,7 +67,7 @@ Using pure cuda.core APIs
============================================================
Device: NVIDIA GeForce RTX XXXX
Kernel compiled ✓
Kernel compiled [OK]
Problem size: 16,000,000 elements (61 MB)

2
python/2_CoreConcepts/streamingCopyComputeOverlap/requirements.txt
View File

@@ -2,5 +2,5 @@
# Requires Python 3.10 or newer
cuda-python>=13.0.0
cuda-core>=0.6.0
cuda-core>=1.0.0
numpy>=2.3.2

22
python/2_CoreConcepts/streamingCopyComputeOverlap/streamingCopyComputeOverlap.py
View File

@@ -102,7 +102,7 @@ def main():
VECTOR_SCALE_KERNEL, code_type="c++", options=ProgramOptions(arch=arch)
)
kernel = program.compile(target_type="cubin").get_kernel("vector_scale")
print("Kernel compiled ✓")
print("Kernel compiled [OK]")
# Parameters
N = 16_000_000 # 16M elements
@@ -126,10 +126,10 @@ def main():
h_in = h_out = d_in = d_out = None
try:
# Pre-allocate buffers
h_in = pinned_mr.allocate(n_bytes, default_stream)
h_out = pinned_mr.allocate(n_bytes, default_stream)
d_in = device_mr.allocate(n_bytes, default_stream)
d_out = device_mr.allocate(n_bytes, default_stream)
h_in = pinned_mr.allocate(n_bytes, stream=default_stream)
h_out = pinned_mr.allocate(n_bytes, stream=default_stream)
d_in = device_mr.allocate(n_bytes, stream=default_stream)
d_out = device_mr.allocate(n_bytes, stream=default_stream)
# Sync before numpy access (numpy operations aren't stream ordered)
default_stream.sync()
@@ -138,7 +138,7 @@ def main():
np_in[:] = np.random.rand(N).astype(np.float32) * 100
config = LaunchConfig(grid=((N + 255) // 256,), block=(256,))
event_opts = EventOptions(enable_timing=True)
event_opts = EventOptions(timing_enabled=True)
# Warm up
h_in.copy_to(d_in, stream=default_stream)
@@ -214,10 +214,10 @@ def main():
h_ins, h_outs, d_ins, d_outs = [], [], [], []
try:
for i in range(n_streams):
h_ins.append(pinned_mr.allocate(chunk_bytes, streams[i]))
h_outs.append(pinned_mr.allocate(chunk_bytes, streams[i]))
d_ins.append(device_mr.allocate(chunk_bytes, streams[i]))
d_outs.append(device_mr.allocate(chunk_bytes, streams[i]))
h_ins.append(pinned_mr.allocate(chunk_bytes, stream=streams[i]))
h_outs.append(pinned_mr.allocate(chunk_bytes, stream=streams[i]))
d_ins.append(device_mr.allocate(chunk_bytes, stream=streams[i]))
d_outs.append(device_mr.allocate(chunk_bytes, stream=streams[i]))
# Initialize input data
for i in range(n_streams):
@@ -245,7 +245,7 @@ def main():
# Benchmark with CUDA events (use stream 0 for timing)
times = []
event_opts = EventOptions(enable_timing=True)
event_opts = EventOptions(timing_enabled=True)
for _ in range(n_runs):
start_ev = device.create_event(options=event_opts)
end_ev = device.create_event(options=event_opts)