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

This is the release of the CUDA 13.3 samples, which include additions for CUDA Tile C++, and updated CCCL and Python samples.

python/2_CoreConcepts/matrixMulSharedMem/README.md

@@ -1,5 +1,25 @@
 # Matrix Multiplication with Shared Memory (GEMM)
 
+> **Known issue — version-pinned sample.** Unlike the other samples in this
+> repository, this sample is pinned to `cuda-core==0.7.0` and
+> `nvmath-python[cu13]==0.9.0`. The reason is that nvmath-python 0.9.0
+> still uses `cuda-core`'s pre-1.0 API name `EventOptions(enable_timing=...)`
+> in its own internals, which `cuda-core>=1.0` no longer accepts.
+>
+> If you install this sample's `requirements.txt` into the same environment
+> as the other samples, pip will downgrade `cuda-core` and the other
+> samples (which use the 1.0 API) will stop working. The recommended
+> workflow is one of:
+>
+> - Install this sample's requirements in a **dedicated virtual
+>   environment**, or
+> - Re-run the other samples' `pip install -r requirements.txt` afterwards
+>   to upgrade `cuda-core` back to 1.0.
+>
+> This sample will be re-aligned with the rest of the repository
+> (`cuda-core>=1.0.0`) once nvmath-python ships a release that targets
+> cuda-core's 1.0 naming audit.
+
 Demonstrates efficient matrix multiplication using nvmath-python APIs and custom CUDA kernels with tiling, shared memory, and loop unrolling.
 
 ## Overview
@@ -78,7 +98,7 @@ Using nvmath and cuda.core APIs
 Device: NVIDIA GeForce RTX 4090
 Compute Capability: sm_89
 
-Custom kernel compiled ✓
+Custom kernel compiled [OK]
 
 Matrix dimensions: A(1024x1024) × B(1024x1024) = C(1024x1024)
 Custom kernel tile size: 16x16

python/2_CoreConcepts/matrixMulSharedMem/matrixMulSharedMem.py

@@ -31,7 +31,7 @@ Demonstrates efficient matrix multiplication using:
 - nvmath.linalg.advanced.Matmul for high-performance GEMM via cuBLASLt
 - Custom CUDA kernel with tiling, shared memory, and loop unrolling
 
-Uses cuda.core APIs with CuPy arrays via ExternalStream.
+Uses cuda.core APIs with CuPy arrays via Stream.from_external.
 """
 
 import sys
@@ -121,17 +121,22 @@ def run_matmul_benchmark(
     print(f"Compute Capability: sm_{device.arch}")
 
     # Make CuPy use our cuda.core stream
-    cp.cuda.ExternalStream(int(stream.handle)).use()
+    cp.cuda.Stream.from_external(stream).use()
 
     # Compile custom kernel
     arch = f"sm_{device.arch}"
     program = Program(MATMUL_KERNEL, code_type="c++", options=ProgramOptions(arch=arch))
     kernel = program.compile(target_type="cubin").get_kernel("matmul_shared")
-    print("Custom kernel compiled ✓")
+    print("Custom kernel compiled [OK]")
 
     # Setup
     print(f"\nMatrix: A({m}x{k}) × B({k}x{n}) = C({m}x{n})")
     total_ops = 2 * m * n * k
+    # NOTE: this sample is pinned to cuda-core==0.7.0 (see requirements.txt)
+    # because nvmath-python 0.9.0 still uses cuda-core's pre-1.0 API name
+    # `enable_timing`. Once nvmath ships a release compatible with cuda-core
+    # 1.0, bump the pins in requirements.txt and rename this kwarg to
+    # `timing_enabled` to match the rest of the samples.
     event_opts = EventOptions(enable_timing=True)
 
     # Allocate matrices

python/2_CoreConcepts/matrixMulSharedMem/requirements.txt

@@ -1,7 +1,20 @@
 # Matrix Multiplication with Shared Memory (GEMM) Requirements
+#
+# IMPORTANT: this sample pins older versions of cuda-core and nvmath-python
+# on purpose. nvmath-python 0.9.0 (the current CUDA-13 release at the time
+# of CTK 13.3) calls cuda-core's pre-1.0 API name `EventOptions(enable_timing=...)`
+# in its own internals. With cuda-core 1.0+ that kwarg was renamed to
+# `timing_enabled` and the old name is rejected, so any cuda-core>=1.0 +
+# nvmath-python 0.9.0 combination raises a TypeError at runtime.
+#
+# Until nvmath-python ships a release that targets the cuda-core 1.0 naming
+# audit, this sample requires the older cuda-core 0.7 line. Installing this
+# requirements.txt into the same environment as the other samples will
+# downgrade cuda-core; use a dedicated venv for this sample, or reinstall
+# the other samples' requirements afterwards to upgrade cuda-core back.
 
 cuda-python>=13.0.0
-cuda-core>=0.6.0
-cupy-cuda13x>=13.0.0
+cuda-core==0.7.0
+cupy-cuda13x>=14.0.0
 numpy>=2.3.2
-nvmath-python[cu13]>=0.3.0
+nvmath-python[cu13]==0.9.0