fullwave25 1.4.0__tar.gz → 1.4.1.dev0__tar.gz

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/.bumpversion.toml

@@ -1,5 +1,5 @@
 [tool.bumpversion]
-current_version = "1.4.0"
+current_version = "1.4.1-dev0"
 parse = """(?x)
     (?P<major>0|[1-9]\\d*)\\.
     (?P<minor>0|[1-9]\\d*)\\.

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: fullwave25
-Version: 1.4.0
+Version: 1.4.1-dev0
 Summary: Fullwave 2.5: Ultrasound wave propagation simulation with heterogeneous power law attenuation modelling capabilities
 Author: Masashi Sode, Gianmarco Pinton
 Maintainer: Masashi Sode

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/fullwave/__init__.py

@@ -66,7 +66,7 @@ try:
     __version__ = version("fullwave")
 except PackageNotFoundError:
     # Update via bump-my-version, not manually
-    __version__ = "1.4.0"
+    __version__ = "1.4.1-dev0"
 
 VERSION = __version__  # for convenience
 logger.info("Fullwave version: %s", __version__)

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/fullwave/solver/binary_manager.py

@@ -28,7 +28,7 @@ CACHE_DIR = Path.home() / ".cache" / "fullwave25" / "bins"
 
 # Pinned release tag for the solver binaries.
 # Update this only when new binaries are uploaded to a GitHub release.
-BINARY_RELEASE_TAG = "fullwave_bin_v1.5"
+BINARY_RELEASE_TAG = "fullwave_bin_v1.5.1"
 
 
 def _download_url(filename: str, tag: str) -> str:

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/fullwave/solver/cuda_utils.py

@@ -213,6 +213,42 @@ def cuCtxDetach(ctx):  # noqa: ANN001, ANN201, D103, N802
     return cuda.cuCtxDetach(ctx)
 
 
+@cuda_api_call
+def cuDeviceCanAccessPeer(can_access, dev, peer_dev):  # noqa: ANN001, ANN201, D103, N802
+    return cuda.cuDeviceCanAccessPeer(can_access, dev, peer_dev)
+
+
+def find_consecutive_gpus_without_peer_access(ordinals: list[int]) -> list[tuple[int, int]]:
+    """Return the consecutive GPU pairs that cannot access each other's memory directly.
+
+    Parameters
+    ----------
+    ordinals : list[int]
+        CUDA device ordinals as this process sees them, in the requested order.
+
+    Returns
+    -------
+    list[tuple[int, int]]
+        Each consecutive pair (a, b) of ``ordinals`` where a cannot access b or b
+        cannot access a.
+
+    """
+    cuInit(0)
+    pairs = []
+    for first, second in zip(ordinals, ordinals[1:], strict=False):
+        device_first = ctypes.c_int()
+        device_second = ctypes.c_int()
+        cuDeviceGet(ctypes.byref(device_first), first)
+        cuDeviceGet(ctypes.byref(device_second), second)
+        forward = ctypes.c_int()
+        backward = ctypes.c_int()
+        cuDeviceCanAccessPeer(ctypes.byref(forward), device_first, device_second)
+        cuDeviceCanAccessPeer(ctypes.byref(backward), device_second, device_first)
+        if not forward.value or not backward.value:
+            pairs.append((first, second))
+    return pairs
+
+
 # Main function to get CUDA device specs
 @lru_cache(maxsize=1)
 def get_cuda_device_specs() -> list[dict[str, Any]]:

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/fullwave/solver/launcher.py

@@ -1,5 +1,6 @@
 """Module for launching Fullwave simulation."""
 
+import ctypes
 import logging
 import os
 import re
@@ -79,6 +80,8 @@ class Launcher:
             cuda_device_id,
             verify_gpu=verify_gpu,
         )
+        if use_gpu and verify_gpu:
+            self._warn_if_gpus_lack_peer_access(self.cuda_device_id)
         self.save_gpu_memory = save_gpu_memory
         logger.debug("Launcher instance created.")
 
@@ -160,6 +163,49 @@ class Launcher:
                 message = f"CUDA device ID {device_id} does not exist."
                 raise ValueError(message)
 
+    @staticmethod
+    def _warn_if_gpus_lack_peer_access(device_id_str: str) -> None:
+        """Warn when consecutive GPUs of a multi-GPU run have no peer-to-peer access.
+
+        Without peer access (no NVLink, and a PCIe chipset or GPU that refuses it,
+        as on GeForce cards), a multi-GPU run is much slower than with peer access.
+
+        The binary runs with ``CUDA_VISIBLE_DEVICES`` set to ``device_id_str``. The
+        CUDA ordinals of this process name the same GPUs only when this process has
+        no ``CUDA_VISIBLE_DEVICES`` of its own, so the check is skipped otherwise.
+        It never raises: if the CUDA driver cannot answer, it logs at debug level.
+
+        Parameters
+        ----------
+        device_id_str : str
+            The CUDA device ID(s) in string format, for example "0,1,2,3".
+
+        """
+        device_ids = [int(device_id) for device_id in device_id_str.split(",")]
+        if len(device_ids) < 2:
+            return
+        if "CUDA_VISIBLE_DEVICES" in os.environ:
+            logger.debug(
+                "Peer access check skipped: CUDA_VISIBLE_DEVICES is set in this process.",
+            )
+            return
+        try:
+            from .cuda_utils import find_consecutive_gpus_without_peer_access  # noqa: PLC0415
+
+            pairs = find_consecutive_gpus_without_peer_access(device_ids)
+        except (AttributeError, NameError, OSError, RuntimeError, ctypes.ArgumentError) as error:
+            logger.debug("Peer access check skipped: %s", error)
+            return
+        if not pairs:
+            return
+        pair_text = ", ".join(f"GPU {first} <-> GPU {second}" for first, second in pairs)
+        logger.warning(
+            "No peer-to-peer access between %s. A multi-GPU run on these GPUs is much "
+            "slower than on GPUs with NVLink or PCIe peer access. Check with a short "
+            "run whether fewer GPUs give a faster simulation.",
+            pair_text,
+        )
+
     @staticmethod
     def _configure_cuda_device_id(
         cuda_device_id: str | int | list | None,

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "fullwave25"
-version = "1.4.0" # Update via bump-my-version, not manually
+version = "1.4.1-dev0" # Update via bump-my-version, not manually
 description = "Fullwave 2.5: Ultrasound wave propagation simulation with heterogeneous power law attenuation modelling capabilities"
 readme = "README.md"
 requires-python = ">=3.10"

{fullwave25-1.4.0 → fullwave25-1.4.1.dev0}/uv.lock

@@ -761,7 +761,7 @@ wheels = [
 
 [[package]]
 name = "fullwave25"
-version = "1.4.0"
+version = "1.4.1.dev0"
 source = { editable = "." }
 dependencies = [
     { name = "cupy-cuda12x" },